USENIX Security '26 Technical Sessions

Papers and Proceedings

The full Proceedings published by USENIX for the symposium are available for download below. Individual papers can also be downloaded from their respective presentation pages. Copyright to the individual works is retained by the author[s].

Proceedings Front Matter
Proceedings Cover | Title Page and List of Organizers | Message from the Program Co-Chairs | Table of Contents

Full Proceedings PDFs
 USENIX Security '26 Full Proceedings (PDF, 809.9 MB)
 USENIX Security '26 Proceedings Interior (PDF, 873.2 MB, best for mobile devices)
 USENIX Security '26 Errata #1 (PDF)
 USENIX Security '26 Errata #2 (PDF)
 USENIX Security '26 Errata #3 (PDF)
 USENIX Security '26 Full Artifact Appendices Proceedings (PDF, 14.87 MB)
 USENIX Security '26 Artifact Appendices Proceedings Interior (PDF, 24.25 MB, best for mobile devices)

Additional Materials
Paper Awards: Runners-Up

Attendee Files 
USENIX Security '26 Attendee List (PDF)

Wednesday, August 12

7:00 am–8:45 am

Continental Breakfast

Grand Foyer North and West

8:45 am–9:30 am

Opening Remarks and Awards

Grand Ballroom V–VI

Program Co-Chairs: Elissa Redmiles, Georgetown University; Ben Stock, CISPA Helmholtz Center for Information Security

9:30 am–10:00 am

Coffee and Tea Break

Grand Foyer North and West

10:00 am–11:30 am

Track 1

Applications of Cryptography 1

Session Chair: Nirvan Tyagi, University of Washington

Grand Ballroom VII-X

Ajax: Fast Threshold Fully Homomorphic Encryption without Noise Flooding

Zhenkai Hu, Shanghai Jiao Tong University and State Key Laboratory of Cryptology; Haofei Liang, Shanghai Jiao Tong University; Xiao Wang, Northwestern University; Xiang Xie, East China Normal University and Primus Labs; Kang Yang, State Key Laboratory of Cryptology; Yu Yu, Shanghai Jiao Tong University and Shanghai Qi Zhi Institute; Wenhao Zhang, Northwestern University

Available Media

Threshold fully homomorphic encryption (ThFHE) enables multiple parties to perform arbitrary computation over encrypted data, while the secret key is distributed across the parties. The main task of designing ThFHE is to construct threshold key-generation and decryption protocols for FHE schemes. Among existing FHE schemes, FHEW-like cryptosystems enjoy the advantage of fast bootstrapping and small parameters. However, known ThFHE solutions use the "noise-flooding" technique to realize threshold decryption, which requires either large parameters or switching to a scheme with large parameters via bootstrapping, leading to a slow decryption process. Besides, for key generation, existing ThFHE schemes either assume a generic MPC or a trusted setup, or incur noise growth that is linear in the number n of parties.

In this paper, we propose a fast ThFHE scheme Ajax, by designing threshold key-generation and decryption protocols for FHEW-like cryptosystems. In particular, for threshold decryption, we eliminate the need for noise flooding, and instead present a new technique called "mask-then-open" based on random double sharings over different rings, while keeping the advantage of small parameters. For threshold key generation, we show a simple approach to reduce the noise growth from n times to max(0.038n,2) times in the honest-majority setting, where at most t=(n-1)/2 parties are corrupted. Our end-to-end implementation reports the running time 17.6 s and 0.9 ms (resp., 91.9 s and 4.4 ms) of generating a set of keys and decrypting a single ciphertext respectively, for n=3 (resp., n=21) parties under the network of 1 Gbps bandwidth and 1 ms ping time. Compared to the state-of-the-art implementation, our protocol improves the end-to-end performance of the threshold decryption protocol by a factor of at least 5.7× 283.6× across different network latencies from t=1 to t=13. Our approaches can also be applied in other types of FHE schemes like BGV, BFV, and CKKS.

BatchBoot: Fast Batched Bootstrapping for TFHE scheme and Practical Applications

Zhihao Li, Ant Digital Technologies, Ant Group; Hongyu Wang, Shanxi University; Yuan Zhao and Lichun Li, Ant Digital Technologies, Ant Group; Zhiwei Wang, State Key Laboratory of Cyberspace Security Defense, Institute of Information Engineering, CAS; Jiaxing He, Changzheng Wei, and Ying Yan, Ant Digital Technologies, Ant Group; Lifeng Guo, Shanxi University

Available Media

Torus-based Fully Homomorphic Encryption (TFHE) is distinguished by its unique bootstrapping mechanism, which enables arbitrary computation while refreshing the noise budget. However, this mechanism exhibits limited scalability since it can handle only a single encrypted message at a time. To address this, recent studies have proposed batched bootstrapping schemes that allow TFHE to process ciphertexts in parallel, thereby achieving promising amortization benefits. Despite these advances, this emerging direction remains underexplored, leaving ample room for further investigation.

In this paper, we present BatchBoot, an efficient batched bootstrapping framework for TFHE that enables amortized processing of encrypted messages. Specifically, our work makes three key contributions. First, we redesign the core submodule, i.e., homomorphic polynomial multiplication, to substantially reduce the reliance on expensive FFT operations. Second, we propose a sparsity-aware message packing strategy that flexibly supports varying packing scales. Third, we extend functional bootstrapping to circuit bootstrapping, thereby greatly enhancing the expressiveness of supported functions. Together, these contributions enable BatchBoot to deliver a 2.4× speedup over the state-of-the-art batched scheme (Guimarães et al., CCS'25) and a 43.8× improvement over the non-batched TFHE-rs implementation.

At the application level, we highlight the versatility of BatchBoot through two practical use cases. First, we present the first TFHE-based PSI protocol under the unbalanced setting, which achieves a 294× reduction in communication cost and a 4.1× speedup compared to the best BFV-based solution (PEPSI, USENIX Security'24). Second, we design an 8-bit FHE instruction set based on the BatchCBoot that delivers up to a 5.4× speedup over the existing results (Wang et al., CCS'25).

HasteBoots: Proving TFHE Programmable Bootstrapping in Seconds

Fengrun Liu, Carnegie Mellon University; Haofei Liang, Shanghai Jiao Tong University; Xiang Xie, East China Normal University and Primus Labs; Yu Yu, Shanghai Jiao Tong University; Wenting Zheng, Carnegie Mellon University; Yuncong Hu, Shanghai Jiao Tong University

Available Media

Fully Homomorphic Encryption (FHE) enables computations on encrypted data, ensuring privacy for outsourced computation. However, verifying the integrity of FHE computations remains a significant challenge, especially for bootstrapping, the most computationally intensive operation in FHE. Prior approaches, including zkVM-based solutions and general-purpose SNARKs, suffer from inefficiencies, with proof generation time ranging from several hours to days.

In this work, we propose HasteBoots, a succinct argument tailored for TFHE with programmable bootstrapping. By designing efficient protocols for arithmetic operations over quotient rings, HasteBoots achieves proof generation in a few seconds for TFHE evaluation with programmable bootstrapping, significantly outperforming the state-of-the-art, Zama (CCS'25). Moreover, HasteBoots supports batching multiple TFHE evaluations and bootstrappings, a feature that prior work cannot practically support due to prohibitive proving cost. HasteBoots can prove a batch of 16 operations within one minute, while maintaining succinct verification, requiring only 126 ms and a proof size of 0.28 MB. Our approach demonstrates the potential for scalable and efficient verifiable FHE, paving the way for practical, privacy-preserving computations.

Orbit: Optimizing Rescale and Bootstrap Placement with Integer Linear Programming Techniques for Secure Inference

Zikai Zhou, Tsinghua University; William Seo and Edward Chen, Carnegie Mellon University; Alex Ozdemir, Max Planck Institute for Security and Privacy; Fraser Brown and Wenting Zheng, Carnegie Mellon University

Available Media

Fully Homomorphic Encryption (FHE) allows computation on encrypted data without decrypting it. In theory, FHE makes privacy-preserving machine learning possible. In practice, however, it remains impractically slow for real workloads. A major source of slowdown is bootstrap operations; in CKKS, a popular FHE scheme for tensor workloads, the slowdown is compounded by scale management and rescale operations.

FHE compilers for machine learning inference aim to make bootstrap placement and scale management efficient and easy by compiling high-level tensor programs into optimized CKKS computations. Unfortunately, existing approaches miss crucial optimization opportunities because they overlook a key property of CKKS programs: bootstrap and rescale placement are fundamentally coupled through the level budget. In this paper, we present Orbit, an FHE compiler that jointly optimizes bootstrap and rescale placement through a novel Integer Linear Programming (ILP) formulation that reasons about both ciphertext level and scale constraints. To make this formulation tractable for structured tensor workloads, particularly convolutional neural networks, we introduce three techniques that reduce ILP complexity while preserving optimality. Across five workloads and multiple cryptographic parameter configurations, Orbit achieves a geometric mean speedup of 19% over DaCapo, 73% over Orion, and 52% over ReSBM, keeps compilation under 6 minutes, and retains model accuracy within 0.3% of plaintext execution.

Libra: Pattern-Scheduling Co-Optimization for Cross-Scheme FHE Code Generation over GPGPU

Song Bian, Yintai Sun, Zian Zhao, and Haowen Pan, Beihang University; Mingzhe Zhang, unaffiliated; Zhenyu Guan, Beihang University

Available Media

We propose Libra, a compiler framework that automates efficient code generation for cross-scheme fully homomorphic encryption (FHE) on highly parallel computing architectures. While it is known that leveraging multiple FHE schemes in a single application can improve the overall efficiency, the exact mapping of cross-scheme FHE operators onto high-performance architectures, such as general-purpose graphic processing units (GPGPUs), remains challenging. To address such challenge, Libra integrates both the FHE computational patterns and hardware-aware scheduling strategies to establish an algorithm-hardware co-optimization framework. Specifically, Libra defines a novel cross-scheme representation for FHE that abstracts common program patterns for each of the FHE schemes. Then, we dynamically optimize the output FHE program based on the combined execution costs of FHE primitives derived from multiple scheme switching patterns. Next, to accelerate inter-operator execution on GPUs, Libra introduces a computational scheduling strategy that bridges high-level computation characteristics with low-level execution plans. Through the proposed pattern-scheduling co-optimization process, Libra generates efficient codes for cross-scheme high-precision FHE computations on GPGPUs. Experiment results show that Libra achieves up to 270× speedup on microbenchmarks and 19× on the applications compared to state-of-the-art cross-scheme, while improving compute unit and memory bandwidth utilization by 44% and 36.1%.

VROOM: Accelerating (Almost All) Number-Theoretic Cryptography Using Vectorization and the Residue Number System

Simon Langowski, Kaiwen He, and Srinivas Devadas, MIT

Available Media

Modular arithmetic with a large prime modulus is a dominant computational cost in number-theoretic cryptography. Modular operations are especially challenging to parallelize efficiently on CPUs using vector instructions. Standard CPU implementations rely on costly carry operations and permutation instructions to align with the multiplication datapath, negating the benefits of vectorization.

We develop vectorized algorithms for modular addition and multiplication, and present a new, constant-time modular multiplication algorithm suitable for general moduli—prime or otherwise. Our method uses a Residue Number System (RNS) representation to align the arithmetic naturally with wide vector units, and strategically eliminate extraneous instructions. Existing works either require the use of customized hardware or fail to show latency improvements.

Reducing the latency of modular arithmetic results in speedups for cryptographic applications. We accelerate RSA-4096 signatures by 4.0× (verify) and 1.3× (sign) over OpenSSL, and speed up BLS signature verifications by 4.05× over the assembly-optimized blst library. To facilitate broad practical adoption, we are upstreaming our implementation into BoringSSL, where it will directly benefit real-world TLS and cryptographic deployments. Results on mapping our algorithm to Nvidia GPUs demonstrate speedups on modular multiplication over Nvidia's CGBN library.

RBOOT: Accelerating Homomorphic Neural Network Inference by Fusing ReLU within Bootstrapping

Zhaomin Yang, Chao Niu, Benqiang Wei, Zhicong Huang, Cheng Hong, and Tao Wei, Ant Group

Available Media

A major bottleneck in secure neural network inference using Fully Homomorphic Encryption (FHE) is the evaluation of non-linear activation functions like ReLU, which are inefficient to compute under FHE. State-of-the-art solutions approximate ReLU using high-degree polynomials, incurring significant computational overhead. We present RBOOT, an optimized framework that seamlessly integrates ReLU evaluation into CKKS bootstrapping, significantly reducing multiplication depth and boosting efficiency. Our key insight is that the EvalMod step in CKKS bootstrapping is composed of trigonometric functions, which are nonlinear themselves. Prior works treat bootstrapping and activation functions as independent routines, missing an opportunity to leverage such nonlinearity. By co-optimizing these components, we can exploit such nonlinearity to construct ReLU (and other non-linear functions) within the bootstrapping process itself, greatly reducing the computation overhead. Results on four widely used CNN models show that RBOOT achieves 2.77× faster end-to-end inference and 81% lower memory usage compared to previous polynomial approximation works, while maintaining comparable accuracy.

Bridging Usability and Performance: A Tensor Compiler for Autovectorizing Homomorphic Encryption

Edward Chen, Fraser Brown, and Wenting Zheng, Carnegie Mellon University

Available Media

Homomorphic encryption (HE) offers strong privacy guarantees by enabling computation over encrypted data. However, the performance of tensor operations in HE is highly sensitive to how the plaintext data is packed into ciphertexts. Large tensor programs introduce numerous possible layout assignments, making it both challenging and tedious for users to manually write efficient HE programs.

In this paper, we present Rotom, a compilation framework that autovectorizes tensor programs into optimized HE programs. Rotom systematically explores a wide range of layout assignments, applies state-of-the-art optimizations, and automatically generates an equivalent, efficient HE program. At its core, Rotom utilizes a novel, lightweight ApplyRoll layout conversion operator to easily modify the underlying data layouts and unlock new avenues for performance gains. Our evaluation demonstrates Rotom scalably compiles all tensor workloads in under 5 minutes, reduces rotations in hand-tuned protocols by up to 3×, and achieves up to 80× performance improvement over prior autovectorization systems.

Track 2

Security of ML 1

Session Chair: Shengwei An, Virginia Tech

Grand Ballroom VI

The Art of Hide and Seek: Making Pickle-Based Model Supply Chain Poisoning Stealthy Again

Tong Liu and Guozhu Meng, Institute of Information Engineering, Chinese Academy of Sciences; School of Cyber Security, University of Chinese Academy of Sciences; Peng Zhou, Shanghai University; Zizhuang Deng, School of Cyber Science and Technology, Shandong University; State Key Laboratory of Cryptography and Digital Economy Security, Shandong University; Shuaiyin Yao and Kai Chen, Institute of Information Engineering, Chinese Academy of Sciences; School of Cyber Security, University of Chinese Academy of Sciences

Available Media

Pickle deserialization vulnerabilities have persisted throughout Python's history, remaining widely recognized yet unresolved. Due to its ability to transparently save and restore complex objects, many AI/ML frameworks continue to adopt pickle as the model serialization protocol despite its inherent risks. As the open-source model ecosystem grows, model-sharing platforms such as Hugging Face have attracted massive participation, significantly amplifying the real-world impact of pickle exploitation and opening new avenues for model supply chain poisoning. Although several state-of-the-art scanners have been developed to detect poisoned models, their incomplete understanding of the poisoning surface allows attackers to bypass them. In this work, we present the first systematic disclosure of the pickle-based model poisoning surface from both model loading and risky function perspectives. Our research demonstrates how pickle-based model poisoning can remain stealthy and highlights critical gaps in current scanning solutions. On the model loading surface, we identify 22 distinct pickle-based model loading paths across five foundational AI/ML frameworks, 19 of which are entirely missed by existing scanners. We further develop a bypass technique named Exception-Directed Programming (EDP) and discover 9 EDP instances, 7 of which can bypass all scanners. On the risky function surface, we discover 133 exploitable gadgets, achieving almost a 100% bypass rate. Even against the best-performing scanner, these gadgets maintain an 89% bypass rate. By systematically revealing the pickle-based model poisoning surface, we achieve practical and robust bypasses against real-world scanners. We responsibly disclose our findings to corresponding vendors, receiving acknowledgments and a $12,000 bug bounty.

Fend for Yourself! Backdoor Purification in Federated Graph Learning with an Evolving Knowledge Anchor

Chengcheng Zhu and Yunlong Mao, Nanjing University; Jiale Zhang and Bosen Rao, Yangzhou University; Sheng Zhong, Nanjing University

Available Media

Federated Graph Learning (FedGL) enables collaborative training on decentralized graph data while preserving privacy, yet its distributed nature makes it highly vulnerable to backdoor attacks. These attacks compromise the integrity of the global model by injecting malicious triggers. Existing defenses, however, are often ineffective on complex graph data or rely on a trusted server, creating an architectural conflict with modern privacy-preserving technologies. To overcome these limitations, we propose GBHINDER, a novel and practical trusted-server-free defense framework where each benign participant defends itself. GBHINDER establishes a virtuous cycle: it leverages its own trusted historical knowledge as a benign anchor to purify the downloaded global model, and in turn, selectively incorporates the global model's benign knowledge to progressively evolve the anchor itself. Specifically, this cycle is driven by two key components. A Historical Channel Attention Regularization module uses the anchor to constrain the global model's representations and disrupt backdoor propagation. To resolve the tension between local trust and global collaboration, an Adaptive Momentum Information Update mechanism enables the anchor to safely evolve by dynamically integrating robust global information, ensuring the anchor remains effective with federated iteration. Extensive experiments on several benchmark datasets demonstrate that GBHINDER significantly outperforms state-of-the-art (SOTA) defenses, successfully reducing the backdoor attack success rate to below 10% while preserving high accuracy on the main task.

Lethe: Purifying Backdoored Large Language Models with Knowledge Dilution

Chen Chen, Nanyang Technological University; Yuchen Sun and Jiaxin Gao, Wuhan University; Xueluan Gong, Nanyang Technological University; Qian Wang, Wuhan University; Ziyao Liu, Yongsen Zheng, and Kwok-Yan Lam, Nanyang Technological University

Available Media

Large language models (LLMs) have seen significant advancements, achieving superior performance in various Natural Language Processing (NLP) tasks. However, they remain vulnerable to backdoor attacks, where models behave normally for standard queries but generate harmful responses or unintended output when specific triggers are activated. Existing backdoor defenses either lack comprehensiveness in practice, focusing on narrow trigger settings, detection-only mechanisms, and limited domains, or fail to withstand advanced scenarios like model-editing-based, multi-trigger, and triggerless attacks. In this paper, we present LETHE, a novel method to eliminate backdoor behaviors from LLMs through knowledge dilution using both internal and external mechanisms. Internally, LETHE leverages a lightweight dataset to train a clean model, which is then merged with the backdoored model to neutralize malicious behaviors by diluting the backdoor impact within the model's parametric memory. Externally, LETHE incorporates benign and semantically relevant evidence into the prompt to distract LLM's attention from backdoor features. Experimental results on classification and generation domains across 5 widely used LLMs demonstrate that LETHE outperforms 8 state-of-the-art defense baselines against 8 backdoor attacks. LETHE reduces the attack success rate of advanced backdoor attacks by up to 98% while maintaining model utility. Furthermore, LETHE has proven to be cost-efficient and robust against adaptive backdoor attacks. The code is provided at https://github.com/Xxxxsir/Lethe. Disclaimer: This paper contains potentially offensive content.

Patcher: Post-Hoc Patching of Backdoored Large Language Models

Anjun Gao, University of Louisville; Yueyang Quan, University of North Texas; Yufei Xia, University of Louisville; Zhuqing Liu, University of North Texas; Minghong Fang, University of Louisville

Available Media

Large language models remain vulnerable to jailbreak backdoor attacks, where adversaries poison safety alignment data to embed hidden triggers that bypass safety mechanisms. Existing defenses often require comprehensive attack information or multiple triggered examples, making them impractical when defenders only observe a single reported failure case without knowing whether it stems from a backdoor attack or a natural alignment bug. This paper presents Patcher, a post-hoc defense framework that repairs backdoored language models using only a single reported failure case and the model parameters. Patcher operates in two stages. First, it localizes backdoor triggers by computing response-conditioned gradient-based saliency scores and applying adaptive clustering to separate triggers from benign context. Second, it patches the model through a constrained fine-tuning objective that breaks the trigger-response association while preserving benign-task utility and robustness to non-triggered jailbreak attacks through KL-divergence constraints. We conduct extensive evaluations across multiple backdoor attack strategies and demonstrate that Patcher successfully localizes triggers and neutralizes backdoors while maintaining model utility. We further show robustness against adaptive attacks designed to evade our defense. This work represents a significant step toward practical defenses against training-time attacks in deployed language models.

When the Aggregator Cheats: Data-Free Backdoors in Federated LLM-based QA Systems

Chenqing Zhu, Southeast University; Yanbo Dai, The Hong Kong University of Science and Technology; Yulong Tian, Nanjing University of Aeronautics and Astronautics; Qingming Li, Zhejiang University; Songze Li, Southeast University and Engineering Research Center of Blockchain Application, Supervision and Management (Southeast University), Ministry of Education

Available Media

Large Language Model (LLM)-based question-answering (QA) systems are increasingly deployed in sensitive domains such as healthcare, mental health counseling, and legal consultation. Federated learning (FL) enables collaborative training without sharing raw client data, for which locally trained models are aggregated at a central server (i.e., a cloud service provider) to obtain a global model. In this paper, we explore the potential vulnerability where a malicious aggregator, who may collude with a third-party vendor, stealthily implants advertisement-type backdoors into federated QA models, without ever accessing client data. The attacker's goals are twofold: (1) preserve clean QA fidelity (i.e., the poisoned model behaves like a clean model on non-triggered queries); and (2) generate highly natural, contextually relevant responses with target advertisements when a trigger appears. Achieving these two goals simultaneously is highly challenging, as naive backdoor injection without knowledge about private data may degrade model's clean performance or fail to inject the target. Motivated by this, we propose to leverage clients' uploaded gradients during training, and develop a two-stage framework for data-free and stealthy poisoning: (1) recover representative training samples from client gradients, and (2) construct poisoning datasets utilizing recovered samples and trigger phrases to inject backdoors into the global model. Experiments across representative QA datasets and LLM families under full fine-tuning and LoRA settings demonstrate that, our method achieves nearly 100% Attack Success Rate (ASR) while incurring negligible degradation on clean tasks. Crucially, reconstructing only 5–20% of gradients suffices to mount a reliable attack, exposing a practical blind spot in the pipeline of federated training of QA LLMs.

Unveiling the Pitfalls of Data-Free Backdoor Detection Against Pre-Trained Models

Quan Zhao, Linkang Du, Yuntao Wang, and Zhou Su, Xi'an Jiaotong University; Zheng Li, Shandong University; Xiangshan Gao, Zhejiang University; Yang Zhang, CISPA Helmholtz Center for Information Security

Available Media

Backdoor attacks pose a significant threat to deep learning models, enabling adversaries to manipulate the output through hidden triggers. Recent detection methods aim to identify backdoors without relying on clean samples or assumptions about attacks. Although they report strong performance, these methods are rarely evaluated on pre-trained models. In this paper, we present the first large-scale study of data-free backdoor detection on pre-trained models. Our benchmark includes more than 30,000 models and covers common backdoor attacks. We find that existing data-free methods fail on most pre-trained models, leading to a false sense of security. Despite our effective improvements, serious vulnerabilities remain.

To address this, we propose using convergence speed as a new side-channel signal for backdoor detection. Using this signal, we reveal the cause of the remaining vulnerabilities and build a novel data-free detector that achieves state-of-the-art performance against existing methods. We further analyze how backdoor attacks evade detection and outline unresolved issues. Our results indicate that detecting backdoor attacks requires further exploration. We hope that our work can draw attention to the vulnerabilities in backdoor detection mechanisms for machine learning systems.

Robust Watermarks Meet Backdoored Models: Evading Diffusion Semantic Watermarks via Stealthy Backdoor

Jinyuan Liu, Tsinghua University; Tianshuo Cong, Shandong Key Laboratory of Artificial Intelligence Security, Shandong University; Pei Li, Shandong University; Tianrui Wang, Tsinghua University; Xinlei He, Wuhan University; Anyu Wang, State Key Laboratory of Cryptography and Digital Economy Security, Tsinghua University; Zhongguancun Laboratory; Xiaoyun Wang, State Key Laboratory of Cryptography and Digital Economy Security, Tsinghua University; Zhongguancun Laboratory; Shandong Institute of Blockchain; National Financial Cryptography Research Center

Available Media

Although semantic watermarking is considered a promising safeguard for images generated by Latent Diffusion Models (LDMs), the reliance of the watermark detection pipeline on neural networks introduces a critical yet underexplored backdoor attack surface. To systematically study this vulnerability, we propose GhostVAE to plant a stealthy backdoor into the encoder of Variational Autoencoder (VAE), enabling reliable evasion of watermark detection. GhostVAE operates in two stages: it first constructs a universal trigger via power spectrum regularization to improve the trigger robustness, and then trains a backdoored VAE encoder with a parameter-aligned objective. Through extensive evaluations across three state-of-the-art semantic watermarking schemes and three widely adopted LDMs, we show that GhostVAE preserves watermark detection performance on benign images (achieving an average true positive rate of 94.4%), while simultaneously enabling highly effective evasion under trigger activation (achieving an average attack success rate of 94.6%). Moreover, we comprehensively analyze seventeen representative defenses and demonstrate that GhostVAE remains stealthy across the input space, parameter space, and latent space. Our work fundamentally undermines the trustworthiness of semantic watermarking systems and highlights that secure deployment of semantic watermarks requires end-to-end security considerations, particularly for neural network components.

Confundo: Learning to Generate Robust Poison for Practical RAG Systems

Haoyang Hu and Zhejun Jiang, The University of Hong Kong; Yueming Lyu, Nanjing University; Junyuan Zhang, The University of Hong Kong; Yi Liu, City University of Hong Kong; Ka-Ho Chow, The University of Hong Kong

Available Media

Retrieval-augmented generation (RAG) is increasingly deployed in real-world applications, where its reference-grounded design makes outputs appear trustworthy. This trust has spurred research on poisoning attacks that craft malicious content, inject it into knowledge sources, and manipulate RAG responses. However, when evaluated in practical RAG systems, existing attacks suffer from severely degraded effectiveness. This gap stems from two overlooked realities: (i) content is often processed before use, which can fragment the poison and weaken its effect, and (ii) users often do not issue the exact queries anticipated during attack design. These factors can lead practitioners to underestimate risks and develop a false sense of security. To better characterize the threat to practical systems, we present Confundo, a learning-to-poison framework that fine-tunes a large language model as a poison generator to achieve high effectiveness, robustness, and stealthiness. Confundo provides a unified framework supporting multiple attack objectives, demonstrated by manipulating factual correctness, inducing biased opinions, and triggering hallucinations. By addressing these overlooked challenges, Confundo consistently outperforms a wide range of purpose-built attacks across datasets and RAG configurations by large margins, even in the presence of defenses. Beyond exposing vulnerabilities, we also present a defensive use case that protects web content from unauthorized incorporation into RAG systems via scraping, with no impact on user experience.

Track 3

Network Security 1

Session Chair: Min Suk Kang, Korea Advanced Institute of Science and Technology (KAIST)

Grand Ballroom V

Analyzing the WebRTC Ecosystem and Breaking Authentication in DTLS-SRTP

Martin Bach, Technology Innovation Institute; Vukašin Karadžić, TU Darmstadt; Lukas Knittel, Ruhr-University Bo­chum; Robert Merget and Jean Paul Degabriele, Technology Innovation Institute

Available Media

DTLS-SRTP was designed to secure real-time media communication and is found in prominent audio and video call platforms, including Zoom, Teams, and Google Meet. Notably, it is part of Web Real-Time Communication (Web-RTC), a web standard enabling real-time communication in the browser. To this end, WebRTC uses multiple technologies, including HTTP, TLS, SDP, ICE, STUN, TURN, UDP, TCP, DTLS, (S)RTP, (S)RTCP, and SCTP. This amalgamation of technologies results in an overly complex system that is very challenging to audit systematically and automatically. As a result, the security of deployments of this core modern communication technology remains largely unexplored.

In this work, we aim to close this gap by developing an automated MitM testing framework (DTLS-MitM-Scanner (DMS)) to test the DTLS channel of a DTLS-SRTP connection. We use our framework to study the current state of the ecosystem in a case study spanning 24 service providers across their browser and mobile applications. Our analysis puts special emphasis on the authentication mechanism in DTLS-SRTP, where we test for 19 potential vulnerabilities that could lead to authentication bypasses for both the client and server. We find that among the 33 tested media server implementations, 19 contained vulnerabilities allowing an attacker to break authentication at the DTLS layer. For 9 of the affected systems, which serve hundreds of millions of users, we could also demonstrate that they could be exploited by an attacker to retrieve media data, assuming only Man-in-the-Middle capabilities. We highlight the impact of these vulnerabilities by building a Proof-of-Concept exploit to listen to Webex video conference calls.

Breaking the Boundaries: Analyzing QUIC Frame-Packet Interactions With QUIC-Attacker

Nurullah Erinola, Marcel Maehren, Marcus Brinkmann, and Jörg Schwenk, Ruhr University Bochum

Available Media

QUIC is a new network protocol based on UDP that replaces TCP and TLS with an integrated protocol. It provides multiplexing of streams over a single encrypted and authenticated connection. The QUIC standard allows many different combinations of UDP datagrams, and QUIC packets, frames, and streams to transport the same information. This implies that testing the receiving side of QUIC is difficult.

We develop probes to explore how different QUIC server implementations handle the coalescence and fragmentation of payloads, covering both valid and invalid combinations of datagrams, packets, and frames. Already at this basic level, we observe significant differences between implementations, some of which pointing towards exploitable vulnerabilities. Previous QUIC research tools were not designed to implement such probes. To address this limitation, we present QUIC-Attacker, a testing framework that allows maximum freedom on the sending side of QUIC.

We present our results on these probes when applied to 15 QUIC server libraries, uncovering eight DoS vulnerabilities caused by unhandled exceptions and exploitable injection vulnerabilities in Kwik and Alibaba's XQUIC.

When Authorization Loses Its Meaning: Breaking and Fixing Third-Party Online Payments

Yongkang Xiao, Jing Chen, Min Shi, Kun He, Qiyi Deng, and Ruiying Du, Wuhan University

Available Media

Third-party online payment systems, such as Alipay and PSPB, constitute critical infrastructure for modern e-commerce. However, their security rests on the unrealistic assumption of fully trusted communication channels. While prior studies have identified isolated vulnerabilities, a systematic formal analysis of payment protocol security remains absent. This paper presents formal security models for six third-party payment protocols, spanning three major payment scenarios and two dominant payment service providers. Our analysis reveals a fundamental design flaw: whenever channel integrity is compromised between the merchant client, merchant server, or payment system, order tampering attacks become feasible. We validate this threat on Android, where over 20% of tested merchant applications allow order tampering through implicit Intent hijacking. To mitigate this threat, we propose user-side order authentication, where per-user-merchant key pairs cryptographically bind consent to order semantics. Formal verification demonstrates its resilience against identified attacks under weak channel assumptions. By bridging formal methods and empirical analysis, this work offers actionable guidance for standardizing secure payment protocols.

WILD Attack: Stealthy Undermining of Wi-Fi-Based Geolocation Through Remote Crowdsourced Data Injection

Changjia Zhu, Xiao Han, Parush Gera, Zhuo Lu, Tempestt Neal, and Yao Liu, University of South Florida

Available Media

Traditional Wi-Fi Positioning System (WPS) spoofing attacks, while seemingly effective, have failed to raise major WPS security concerns due to their lack of stealth and persistence. This paper introduces a novel WILD Attack that undermines WPS security by subverting its core infrastructure–the Location Lookup Table (LLT). In this attack, an adversary remotely submits falsified crowd-sourced reports for target Wi-Fi access points, inducing WPS providers to update LLT based on falsified rather than legitimate data. We examine four widely deployed WPS providers–Google, Apple, A-Map, and WiGLE–and observe that they all accept falsified reports and apply distinct policies to resolve conflicts between legitimate and falsified data. Exploiting these policies, the attacker can induce two forms of LLT subversion: LLT Entry Tampering and LLT Entry Removal, both persisting for weeks even after the attacker ceases activity. We further present three case studies that show the real-world impact of the WILD Attack and propose countermeasures to mitigate such threats.

Opossum Attack: Application Layer Desynchronization using Opportunistic TLS

Robert Merget, Technology Innovation Institute; Nurullah Erinola and Marcel Maehren, Ruhr University Bochum; Lukas Knittel, Ruhr University Bo­chum; Sven Hebrok, Paderborn University; Marcus Brinkmann, Ruhr University Bochum; Juraj Somorovsky, Paderborn University; Jörg Schwenk, Ruhr University Bochum

Distinguished Paper Award Winner

Available Media

Many protocols, like HTTP, FTP, POP3, and SMTP, were originally designed as synchronous plaintext protocols – commands and data are sent in the clear, and the client waits for the response to a pending request before sending the next one. Later, two main solutions were introduced to retrofit these protocols with TLS protection. (1) Implicit TLS: Designate a new, well-known TCP port for each protocol-over-TLS, and start with TLS immediately. (2) Opportunistic TLS: Keep the original well-known port and start with the plaintext protocol, then switch to TLS in response to a command like STARTTLS.

In this work, we present a novel weakness in the way TLS is integrated into popular application layer protocols through implicit and opportunistic TLS. This weakness breaks authentication, even in modern TLS implementations if both implicit TLS and opportunistic TLS are supported at the same time. This authentication flaw can then be utilized to influence the exchanged messages after the TLS handshake from a pure MitM position.In contrast to previous attacks on opportunistic TLS, this attack class does not rely on bugs in the implementations and only requires one of the peers to support opportunistic TLS.

We analyze popular application layer protocols that support opportunistic TLS regarding their vulnerability to the attack. To demonstrate the practical impact of the attack, we analyze exploitation techniques for HTTP (RFC 2817) in detail, and show four different exploit directions. To estimate the impact of the attack on deployed servers, we conducted a series of IPv4-wide scans over multiple protocols and ports to check for support of opportunistic TLS. We found that support for opportunistic TLS is still widespread for many application protocols, with over 3 million servers supporting both, implicit and opportunistic TLS at the same time. In the case of HTTP, we found 20,121 servers that support opportunistic HTTP across 35 ports, with 2,268 of these servers also supporting HTTPS and 539 using the same domain names for implicit HTTPS, presenting an exploitable scenario.

E2E-AKMA: An End-to-End Secure and Privacy-Enhancing AKMA Protocol Against the Anchor Function Compromise

Yueming Li, Institute of Software, Chinese Academy of Sciences; and University of Chinese Academy of Sciences; Long Chen, Institute of Software, Chinese Academy of Sciences; and Key Laboratory of System Software, Chinese Academy of Sciences; Qianwen Gao, Institute of Software, Chinese Academy of Sciences; and University of Chinese Academy of Sciences; Zhenfeng Zhang, Institute of Software, Chinese Academy of Sciences; and Key Laboratory of System Software, Chinese Academy of Sciences

Available Media

The Authentication and Key Management for Applications (AKMA) system represents a recently developed protocol established by 3GPP, which is anticipated to become a pivotal component of the 5G standards. AKMA enables application service providers to delegate user authentication processes to mobile network operators, thereby eliminating the need for these providers to store and manage authentication-related data themselves. This delegation enhances the efficiency of authentication procedures but simultaneously introduces certain security and privacy challenges that warrant thorough analysis and mitigation.

The 5G AKMA service is facilitated by the AKMA Anchor Function (AAnF), which may operate outside the boundaries of the 5G core network. A compromise of the AAnF could potentially allow malicious actors to exploit vulnerabilities, enabling them to monitor user login activities or gain unauthorized access to sensitive communication content. Furthermore, the exposure of the Subscription Permanent Identifier (SUPI) to external Application Functions poses substantial privacy risks, as the SUPI could be utilized to correlate a user's real-world identity with their online activities, thereby undermining user privacy.

To mitigate these vulnerabilities, we propose a novel protocol named E2E-AKMA, which facilitates the establishment of a session key between the User Equipment (UE) and the Application Function (AF) with end-to-end security, even in scenarios where the AAnF has been compromised. Furthermore, the protocol ensures that no entity, aside from the 5G core network, can link account activities to the user's actual identity. This architecture preserves the advantages of the existing AKMA scheme, such as eliminating the need for complex dynamic secret data management and avoiding reliance on specialized hardware (apart from standard SIM cards). Experimental evaluations reveal that the E2E-AKMA framework incurs an overhead of approximately 9.4% in comparison to the original 5G AKMA scheme, which indicates its potential efficiency and practicality for deployment.

Secure Protocol Composition under Dynamic Corruption: Scaling Up Symbolic Analysis for Real-World Security Properties

Cas Cremers, Erik Pallas, and Aleksi Peltonen, CISPA Helmholtz Center for Information Security

Available Media

Although automated symbolic protocol verification has proven valuable and effective, current approaches begin to reach their limits: While small protocols can be analyzed automatically, the most complex case studies often require substantial expert time and resources. There have been many attempts to solve this problem by compositional verification, but they rely on unrealistic protocol assumptions and do not support real-world security properties like Forward Secrecy.

In this work, we enable compositional symbolic analysis for real-world security protocols with respect to modern security properties. We develop a composition result in the Applied π-Calculus that holds even in the presence of attackers capable of dynamic corruption if the protocols satisfy a disjointness requirement.

We demonstrate the applicability and effectiveness of our result on the composition of a data exchange protocol with a Diffie-Hellman key exchange and a compositional analysis of Forward Secrecy in TLS 1.3 within the scope of RFC 8446 and the ECH extension. While monolithic analyses of TLS 1.3 with ECH fail to deliver a result in 10% of cases, all compositional analyses succeed. Additionally, runtime decreases by 71% and memory usage by 86% on average.

Semantics Over Syntax: Uncovering Pre-Authentication 5G Baseband Vulnerabilities

Qiqing Huang and Xingyu Wang, University at Buffalo; Wanda Guo and Guofei Gu, Texas A&M University; Hongxin Hu, University at Buffalo

Available Media

Modern 5G user equipment (UE) processes Radio Resource Control (RRC) configuration messages during early control-plane exchanges, before authentication and integrity protection are established. Prior work for testing 5G UEs has largely focused on constructing syntactically invalid inputs. In contrast, we show that syntactically valid but semantically inconsistent messages, which violate specification-level field constraints or cross-field dependencies, can drive baseband implementations into invalid states, triggering assertion failures or modem crashes. These findings reveal semantic inconsistencies in pre-authentication signaling as a critical yet underexplored attack surface in 5G UE implementations. To address this gap, we present Constraint-Guided Semantic Testing (CONSET), a framework that systematically extracts specification-level constraints and leverages them to generate targeted semantic violations for testing 5G UEs. CONSET decodes RRC messages into structured fields, derives schema-based rules, infers cross-field dependencies using a Large Language Model (LLM) in an evidence-bounded manner, and produces syntactically valid test cases that intentionally violate semantic constraints. We evaluate CONSET on both commercial and open-source 5G UEs. On commercial smartphones, it uncovers 7 previously unknown vulnerabilities through responsible disclosure, including 3 high-severity CVEs, affecting 64 chipset models and over 542 commercially available smartphone models. On the open-source OAI UE, CONSET additionally triggers 46 distinct crash sites.

Track 4

Hardware Security 1

Session Chair: Tao Ni, City University of Hong Kong

Grand Ballroom I-IV

TONTOU: On the Exploitability of Time-of-Neutralization to Time-of-Use Windows

Daniël Trujillo and Mengjia Yan, MIT CSAIL

Available Media

Recently deployed Spectre v2 mitigations neutralize branch predictor state when switching privilege contexts or immediately prior to indirect branch execution, either through domain isolation or sanitization. These defenses assume that subsequent branch predictor behavior remains free from attacker influence until the neutralized state is used.

Unfortunately, this paper shows that this assumption does not hold on recent AMD and Intel CPUs. We find that postneutralization (Time-of-Neutralization to Time-of-Use, TONTOU) windows can be exploited by an attacker to re-poison the predictor. Specifically, within the post-neutralization window, the attacker can re-direct control-flow of the victim to a training gadget that updates the predictor.

To re-direct control-flow, we introduce INTERRUPT INJECTION, a primitive that exploits post-neutralization windows by leveraging the fact that interrupts can occur at nearly any point in time. Using this primitive, we demonstrate that an attacker can trigger mispredictions during kernel execution on recent AMD and Intel CPUs. To prove its practicality, we build an end-to-end exploit using INTERRUPT INJECTION that leaks arbitrary kernel memory on AMD Zen 2 at a rate of 5.47 bytes/s, despite the latest neutralization techniques.

ARM MTE Performance in Practice

Taehyun Noh, University of Texas at Austin; Yingchen Wang, University of California Berkeley; Tal Garfinkel, Google; Mahesh Madhav, Ampere Computing; Daniel Moghimi, Google; Mattan Erez and Shravan Narayan, University of Texas at Austin

Available Media

We present the first comprehensive analysis of ARM MTE hardware performance on four different microarchitectures: ARM Big (A7x), Little (A5x), and Performance (Cortex-X) cores on the Google Pixel 8 and Pixel 9, and on Ampere Computing's AmpereOne CPU core. We also include preliminary analysis of MTE on Apple's M5 chip. We investigate performance in MTE's primary application—probabilistic memory safety—on both SPEC CPU benchmarks and in server workloads such as RocksDB, Nginx, PostgreSQL, and Memcached. While MTE often exhibits modest overheads, we also see performance slowdowns up to 6.64× on certain benchmarks. We identify the microarchitectural cause of these overheads and where they can be addressed in future processors. We then analyze MTE's performance for more specialized security applications such as memory tracing, time-of-check time-of-use prevention, sandboxing, and CFI. In some of these cases, MTE offers significant advantages today, while the benefits for other cases are negligible or will depend on future hardware. Finally, we explore where prior work characterizing MTE performance has either been incomplete or incorrect due to methodological or experimental errors.

Amortizing Randomness Cost: Efficient Masked Implementation of SDitH Signatures with Common Shares

Guowei Liu, Weijia Wang, Lixuan Wu, Chaoran Wang, Yanhong Fan, Jinliang Wang, and Meiqin Wang, Shandong University

Available Media

MPC-in-the-Head (MPCitH) signatures are attractive for postquantum deployment, but they are structurally vulnerable to side-channel leakage: verification opens almost all simulated views, so any leakage correlated with the remaining unopened view(s) can be amplified and exploited to recover the witness. Among the MPCitH-based signature candidates considered in NIST’s additional-signature round, Syndrome-Decodingin-the-Head (SDitH)-Hypercube is particularly susceptible to this concern, since its signing process combines: (i) offline last-party completion involving nonlinear arithmetic computations, (ii) online Baum–Nof (BN) arithmetic checking, and (iii) Keccak sponge calls that generate or absorb secretbearing values.

To the best of our knowledge, this is the first protocol-compatible masked implementation-level protection architecture for SDitH-Hypercube on Cortex-M4. Our implementation combines three techniques, each protecting one attackcritical hotspot: (i) Common Shares to amortize the masking randomness required by masked multiplications in the offline phase, (ii) a Sec/Pub typing discipline that reconstructs transcript-public and verifier-reconstructible values early to avoid costly Sec×Sec gadgets in the online BN kernels, and (iii) selective masked Keccak backends applied only to secretbearing sponge call sites.

On the L1 parameter set, Common Shares reduces arithmetic masking randomness from 8.37 KiB/signature of the Strong Non-Interference (SNI) profile to 16 B/signature, while the Sec/Pub-typed optimization removes 3,264 B/signature from the online BN kernels. When Keccak is masked, masking randomness is dominated by the sponge: 41,118.75 KiB/signature with DOM-Keccak and 82,237.50 KiB/signature with SNI-Keccak, motivating selective masked hashing and tunable backends. We implement a low-stack reference signer and all protected signers on Cortex-M4. Across protected profiles, signing takes 913.7–1,542.0 Mcycles (1.84×–3.10× over the *Corresponding Author. reference), and fits within 155.6 KiB flash, 136.7 KiB static RAM, and 7.96 KiB signing stack. A 1M-trace first-order fixed-vs-random Test Vector Leakage Assessment (TVLA) experiment on Cortex-M4 no longer detects the previously observed leakage peaks after protection.

Principled Design of Indexing Functions for Memory Coloring

Stephan Dübler and Jana Hofmann, Max Planck Institute for Security and Privacy (MPI-SP); Boris Köpf and Stavros Volos, Azure Research, Microsoft

Distinguished Artifact Award Winner

Available Media

On shared multi-core CPUs, memory coloring achieves microarchitectural isolation by partitioning resources between trust domains. Previous work has shown that memory coloring can, in principle, be used to isolate multiple components simultaneously, e.g., both the last-level (L3) cache and DRAM. The number of colors obtained by these methods depends on the algebraic properties of the indexing functions, though, and today's off-the-shelf CPU designs are often not suited for multi-component memory coloring.

We develop algorithms to automatically synthesize indexing functions that guarantee a minimum number of colors. Given a set of typical workloads and some context on the CPU design, our algorithms compute a new linear indexing function that supports the requested number of colors while maximizing the performance on the workloads. Our approach is based on the observation that the number of colors depends on the overlap of the algebraic kernels of the involved indexing functions. Building on this observation, we translate the requirements that the CPU design imposes on the function into algebraic constraints that our algorithms enforce.

In a case study on a 16-core server-class CPU, we show that our framework yields a coloring scheme that partitions the L3 cache and DRAM banks, increasing the number of colors from 1 to 16 while incurring less than 2.5% performance overhead, on average across SPEC and PARSEC benchmarks.

An Integer Overflow Endgame: Compiler and Architecture Support for Default-On Integer Overflow Mitigation

Zheng Zhang, Qi Ling, and Kian Kasad, Purdue University; Brendan Ryan Sweeney, UT Austin; Deepak Gupta, Meta Platforms Inc.; Tal Garfinkel, Google; Kazem Taram, Purdue University

Distinguished Paper Award Winner

Available Media

Integer overflow checks are off-by-default in most production settings due to their high overheads. Consequently, many serious vulnerabilities remain unmitigated. We analyze the source of these overheads---in compilers including LLVM (UBSan, Rust), and in processors including x86-64, ARM64, and RISC-V---and show how to eliminate them.

Overheads in LLVM stem from how overflow semantics are expressed in LLVM IR. LLVM's current approach, based on intrinsics and branches, interferes with a variety of middle end optimizations. We develop an alternative approach that tracks overflow semantics with metadata, but otherwise leaves LLVM IR unchanged, eliminating this overhead. Overheads in hardware primarily result from current ISA designs, that require doubling the number of μops for checked vs. unchecked arithmetic, stressing the processor pipeline. We show how simple ISA extensions can alleviate this stress, enabling overflow-checked arithmetic to issue as a single μop.

We evaluate these extensions through compiler-based emulation on real hardware and simulation in gem5. Our work offers a path to significantly reducing the cost of overflow checks in existing compilers, and demonstrates that integer overflow mitigation can be cheap enough for default-on use in future hardware.

Khost: KVM-based Near Native MCU Firmware Rehosting

Chunlin Wang, Yicheng Yang, Yuan Zhang, Haoyu Xiao, Yifan Zhang, and Jiarun Dai, Fudan University

Available Media

Microcontroller Unit (MCU)-based devices constitute a critical layer of the Internet of Things (IoT) infrastructure, so ensuring their security is of paramount importance. Rehosting-based dynamic MCU firmware analysis is an effective approach to securing these devices. However, existing rehosting frameworks commonly suffer from substantial performance overhead due to emulation or diminished execution scope.

To address these limitations, we propose Khost, a near-native, scope-preserving rehosting framework. It extends the KVM by introducing a lightweight extended CPU, an auxiliary page table, and a software-based interrupt controller, enabling MCU firmware to be rehosted on high-performance platforms with minimum overhead. It also provides a memory-mapped I/O (MMIO) monitor for quick peripheral interactions and a wrapper for firmware to enable coverage collection and configure the existing fuzzing engines flexibly. Evaluations on two standard benchmarks show that Khost reduces overhead by 90.0% to 95.5% for complex computational tasks and by up to 98.5% for MCU system-level operations, compared to QEMU. Furthermore, fuzzing on 12 real-world firmware with Khost achieves up to 197.5× higher throughput and improves basic block coverage by 6x compared to existing fuzzing tools. Additionally, Khost successfully uncovers 5 previously unknown bugs.

HyperAudit: Towards User Transparent and Highly Efficient System Auditing for Cloud Platforms

Renpeng Zhang, Beihang University and Zhongguancun Lab; Kai Shen, Beihang University; Peng Jiang, Southeast University; Ding Li, Peking University and Zhongguancun Lab; Shujiang Wu, Beihang University; Lei Wang, Beihang University and Zhongguancun Lab

Available Media

System auditing is a critical security primitive for cloud platforms, but existing auditing frameworks place the log collection module inside the compromise-prone guest kernel. This design exposes a large attack surface and enables race condition attacks, where an attacker that compromises the kernel can tamper with uncommitted logs and erase pre-compromise traces. Prior secure collectors mitigate this by frequent synchronous or timed submissions, which incur non-trivial runtime overhead.

We present HyperAudit, a hypervisor-assisted auditing architecture that isolates the log collection module from the guest kernel while keeping collection efficient and deployable. HyperAudit injects a kernel-independent collector into a hidden memory region, protects its code with execute-only permissions, and shields the log buffer with guard-page trapping, minimizing the exposed attack surface. Logs are asynchronously pulled by a consumer in a dedicated Secure VM, avoiding periodic synchronous commits. Under the common auditing assumption that logs generated after kernel compromise are untrustworthy, HyperAudit guarantees the integrity of all pre-compromise logs against race condition and flooding attacks. We implement HyperAudit on both x86 and Arm without extra hardware, and show that it reduces log-intensive application overhead by 49% and 66% compared to eAudit and HitchHiker, even outperforming host-only collectors.

Anchors that Don't Lift: Understanding Supply Chain Driven Kernel Lock-In and Governance-Mediated Mitigation Strategies in SOHO Devices

Ritwik Badola, IIT Madras; Rajdeep Ghosh, IIT Kharagpur; Ashita Gupta, IIIT Kottayam; Chester Rebeiro, IIT Madras; Mainack Mondal, IIT Kharagpur

Available Media

Small Office/Home Office (SOHO) devices are widely popular, yet often attacked due to security vulnerabilities in their firmware, affecting tens of thousands of devices at a time. These security vulnerabilities often stem from outdated Linux kernel versions included in SOHO device firmware. Naturally, prior work audited the extent and impact of this issue by simple Linux version extraction and version number based vulnerability mapping. However, it is unclear how many of these anticipated vulnerabilities actually exist in the heavily customized SOHO kernels and if there are any barriers towards updating Linux kernels in SOHO firmwares.

To address this gap, we uncover actual kernel-related vulnerabilities found in 306 SOHO devices using a high-precision template-based CVE detection mechanism on GPL source releases of more than 900 firmwares from these devices (multiple versions per device). Next, as a first, we traced the supply chain of these vulnerable SOHO devices at scale and identify kernel lock-in as a significant security issue---SOHO vendors are effectively locked to specific (often older) kernel versions due to the system-on-chip (SoC) SDKs they use. This kernel lock-in produces a vulnerability debt that is inherited along the supply chain from SoC vendor to firmware creators (ODM/OEM) to router/IP-camera vendor and ultimately borne by end users. All five SoC vendors in our dataset have used SDKs with Linux kernels that had reached End of Life (EoL) more than a year before their usage in a SOHO device. Finally, we explore the mitigation-potential of individual, regulatory and community governance by analyzing social media posts, regulations and community efforts. Our results show that regulation compliance is insufficient and only SoC vendors who engage with communities for kernel upgradation offered a viable path towards mitigation. We conclude by discussing broader implications of our work for improving holistic security of the supply chain for SOHO devices.

Track 5

Enigma Track: Where Security Meets People

Session Chair: David Freeman, Delphi Research

Dover Room

The Six Developers You'll Meet in Security: Why Secure by Design Needs a Behavioral Playbook

Enrique Larios Vargas, OWASP

Available Media

The Secure by Design movement has advanced security by shifting responsibility from individual developers to the systems they use. However, it assumes developers will respond uniformly to secure defaults and guardrails. They don't.

Using the COM-B behavioral model, which analyzes behavior through Capability, Opportunity, and Motivation, and synthesizing empirical research on developer security practices, I identified six developer personas reflecting fundamentally different patterns of security engagement. A developer skeptical about security's ROI requires a different intervention than one who is enthusiastic but lacks organizational support, or one who actively resists security processes.

This talk introduces the SECUR-E framework (Skeptic, Enthusiast, Compliant, Unaware, Resistant, Embedded) and explains why recognizing behavioral diversity is essential for Secure by Design to succeed. When organizations approach developer security as a behavioral design challenge rather than a compliance issue, they can tailor interventions to the actual barriers of each developer type.

Dr. Enrique Larios Vargas is a Security and Learning Specialist at Adyen in the Netherlands, where he designs and leads large-scale security enablement programs for engineering teams in a global fintech environment. His work focuses on integrating behavioral science into secure software development, helping organizations move beyond compliance-driven approaches toward sustainable security culture.

With over five years of experience across fintech, cybersecurity, and higher education, Enrique bridges research and practice. Prior to joining the industry, he served as a university lecturer and conducted research on behavioral factors in software engineering and cybersecurity. He holds a Ph.D. in Computer Science from Leiden University.

Enrique is an active OWASP contributor, co-leads the OWASP Security Culture Project, and regularly speaks on human-centric cybersecurity, developer behavior, and security culture transformation. His mission is to make secure development not just a policy requirement, but the natural and easiest choice for developers and software teams.

Broken Promises: Who's Accountable When VPNs Fail?

Roya Ensafi, University of Michigan

Available Media

Systematic research across five years, encompassing over a thousand users, nine VPN providers, and hundreds of desktop and mobile apps, reveals that VPNs routinely fail the users who depend on them most. For example:

  • Kill switches leak real IP addresses during tunnel failures
  • DNS queries escape the tunnel, exposing browsing activity to ISPs and adversaries,
  • VPN apps transmit cleartext traffic, deliver configuration files unencrypted, and exfiltrate advertising IDs for cross-app tracking.

These failures are not bugs but the predictable consequence of unregulated markets, misaligned incentives, and app stores that substitute self-reported safety claims and paid verification badges for independent auditing.

Roya Ensafi is a Morris Wellman Associate Professor of Computer Science and Engineering at the University of Michigan. She is the founder of Censored Planet lab which focuses on using data-driven approach to detecting and defending against powerful network intermediaries, government threat actors, and technologies that impact users' freedom of expression online. Prof. Ensafi has studied Russia's throttling of Twitter, HTTPS interception in Kazakhstan, and China's Great Cannon attack, among many other instances of network interference. She is a recipient of the Sloan Research Fellowship, NSF CAREER, Google Faculty Research Award, multiple IRTF Applied Networking Research Prizes, and the Consumer Reports Digital Lab Fellowship. Her work has been cited in popular publications such as The New York Times, Newsweek, Business Insider, Wired, and Ars Technica.

Youth Digital Immersion: Cyber Readiness Builds Mental Health Resilience

Vivek Chaudhuri, SafeTeensOnline Global Head of Partnerships

Available Media

Seventy percent of youth report negative online experiences annually, yet most interventions still rely on restriction, approaches that fail to reflect how young people actually navigate digital environments.

Drawing on a systematic review of 52+ studies and survey data from 1,869 students across four countries, I developed a Youth Mental Health Risk Assessment Framework mapping how usage patterns, individual characteristics, and environmental context shape online harm—and how awareness functions as a measurable protective factor. Effective intervention means meeting youth where they are. Approaches that match how young people actually learn, gamification chief among them, are central to translating research into practice.

New developments include a prototype ML model identifying structured risk signals and connections with a gamified cybersecurity learning system built by HackerRangers for school deployment.

Ultimately, individualized digital readiness coupled with innovative learning approaches empowers youth to be capable digital citizens.

Vivek Chaudhuri is a digital entrepreneur, Gen Z change-maker, and Director of Global Partnerships at SafeTeens Online (STO), a youth-led nonprofit advancing cyber-safety education and digital resilience. STO supports a global community of 5M+ teens across five countries and partners with organizations including the U.S. Department of Homeland Security and the National Cybersecurity Alliance. In his role, Vivek leads the research, design, and delivery of youth-focused programming on cyber safety, digital literacy, and online resilience—translating complex online risks and human-factor challenges into practical, actionable learning. He is also expanding STO's impact into underserved communities in Africa, beginning with Ghana. Vivek has previously spoken at the National Cybersecurity Alliance's CONVENE Conference, sharing youth-centered insights with cybersecurity, law enforcement, and policy leaders. As a high school junior, he has also built award-winning digital solutions and enjoys programming, experimenting with generative AI, fencing, and practicing Taekwondo.

11:30 am–1:00 pm

Lunch (on your own)

1:00 pm–2:30 pm

Track 1

Security and Privacy for Web, Mobile, and Emerging Technologies 1

Session Chair: Yan Long, The Hong Kong University of Science and Technology (Guangzhou)

Grand Ballroom VII-X

Sirens' Whisper: Inaudible Near-Ultrasonic Jailbreaks of Speech-Driven LLMs

Zijian Ling, Huazhong University of Science and Technology and Tsinghua University; Pingyi Hu, Xiuyong Gao, and Xiaojing Ma, Huazhong University of Science and Technology; Man Zhou, Huazhong University of Science and Technology and Tsinghua University; Jun Feng and Songfeng Lu, Huazhong University of Science and Technology; Dongmei Zhang and Bin Benjamin Zhu, Microsoft Corporation

Available Media

Speech-driven large language models (LLMs) are increasingly accessed through speech interfaces, introducing new security risks via open acoustic channels. We present Sirens' Whisper (SWhisper), the first practical framework for covert prompt-based attacks against speech-driven LLMs under realistic black-box conditions using commodity hardware. SWhisper enables robust, inaudible delivery of arbitrary target baseband audio—including long and structured prompts—on commodity devices by encoding it into near-ultrasound waveforms that demodulate faithfully after acoustic transmission and microphone nonlinearity. This is achieved through a simple yet effective approach to modeling nonlinear channel characteristics across devices and environments, combined with lightweight channel‑inversion pre‑compensation. Building on this high‑fidelity covert channel, we design a voice‑aware jailbreak generation method that ensures intelligibility, brevity, and transferability under speech-driven interfaces. Experiments across both commercial and open-source speech-driven LLMs demonstrate strong black-box effectiveness. On commercial models, SWhisper achieves up to 0.94 non-refusal (NR) and 0.925 specific-convincing (SC). A controlled user study further shows that the injected jailbreak audio is perceptually indistinguishable from background-only playback for human listeners. Although jailbreaks serve as a case study, the underlying covert acoustic channel enables a broader class of high-fidelity prompt-injection and command-execution attacks.

When HTTP 402 Meets the Blockchain: Risks on Emerging x402 Payments

Qinying Wang, EPFL; Yong Yang, Zhejiang University; Yuan Chen, Independent Researcher; Shouling Ji, Zhejiang University; Mathias Payer, EPFL

Available Media

x402 is an emerging payment protocol for Web APIs and autonomous AI agents. It is driven by the rise of LLM-based agents that can autonomously purchase access to online services. x402 extends HTTP 402 with a payment negotiation flow and delegates payment proof verification and on-chain settlement to third-party facilitators. As a result, facilitators serve as a shared payment infrastructure for many independent merchants. This centralizes trust and validation in one component, so a single flaw can affect many services. Despite rapid adoption by major vendors and economically meaningful mainnet activity, the security posture of real-world x402 deployments remains poorly characterized.

We present the first systematic study of authorization correctness and execution safety in current facilitator-mediated x402 deployments in the wild, identifying eight security rules for facilitators as critical payment infrastructure. Based on our analysis of rule violations, we derive four new attack vectors, including Free Shopping, Asset Theft, Service Denial, and Gas Abuse. These attacks exploit weaknesses in the real-world facilitator and server implementations and cause severe harm, including direct financial loss to merchants, theft of facilitator-held assets, unbounded sponsor-paid gas/fees, and disruption of payment services. To assess the security of x402 deployments at scale, we propose a semi-automated black-box tool and apply it to 15 major x402 facilitators collectively used by over 60K sellers and 360K buyers. Alarmingly, we find violations in all evaluated facilitators. We responsibly disclosed our findings to the affected parties, who acknowledged the issues and adopted mitigations, including changes by Coinbase. Finally, we complement our controlled testing with an empirical measurement of over 119 million recent Base and Solana transactions, quantifying x402 adoption, facilitator centralization, and ecosystem-level risk indicators.

PANGOLIN: Fuzzing Multilingual IoT Firmware with LLM-Driven Code Analysis

Zhipeng Jia and Xiaokang Yin, Information Engineering University; Shuitao Gan, Laboratory for Advanced Computing and Intelligence Engineering; Chao Zhang, Institute for Network Sciences and Cyberspace, Tsinghua University; JCSS, Tsinghua University (INSC) - Science City (Guangzhou) Digital Technology Group Co., Ltd.; Hangtian Liu, State Key Laboratory of Mathematical Engineering and Advanced Computing; Jiangan Ji, Enzhou Song, and Ruijie Cai, Information Engineering University; Jinglei Tan, State Key Laboratory of Mathematical Engineering and Advanced Computing; Shengli Liu, Information Engineering University

Available Media

Multilingual IoT typically refers to the use of multiple languages to implement its web services, such as C, Python, Lua, etc. While some user-accessible interfaces are visualized through the frontend for interaction, a large number of interfaces remain hidden and are not exposed to the frontend in multilingual IoT. Additionally, their parameters often exhibit complex hierarchical structures. Effectively extracting interface specifications from multilingual devices for vulnerability discovery is an urgent problem that remains unresolved. In this paper, we present PANGOLIN, a novel fuzzing solution designed for multilingual IoT devices. First, we utilize LLMs to analyze API dispatching mechanisms and identify interfaces. Then, we introduce an LLM agent to perform cross-language analysis and generate input parameter specifications. Lastly, we utilize response-driven feedback to correct parameter specifications. This knowledge enables semantics-aware fuzzing that can explore deeper code paths and discover more vulnerabilities. PANGOLIN successfully discovered 68 previously unknown vulnerabilities, i.e., 2.96X more than SOTA tool LABRADOR. Notably, 45 of these vulnerabilities were found in hidden interfaces, whereas EAGLEYE was only able to identify 4 such cases. As of the time of writing, all vulnerabilities have been reported to vendors and acknowledged, with 31 vulnerability IDs assigned.

Estimating the Amount of Script-generated Traffic in a Mixture

Cormac Herley, Microsoft Research

Available Media

We address the question of estimating the fraction of traffic that is bot-generated in a mixture. That is, we seek to estimate α when what we receive is α · Clean + (1-α) · Bot. This is primarily of interest when traffic is attempting to masquerade as human-generated (eg, click-fraud, inauthentic social media engagement, etc).

When at least one pair of features is independent in the clean traffic (eg, time-invariance of geographic distribution) we show that getting an upper-bound on α is equivalent to finding the rank-one matrix that maximizes a simple objective function. We give an efficient method for solving, and derive the tightness of the bound. When we have limited data, error analysis is extremely important, since the sampled version of a rank-one matrix will not be precisely rank-one. We derive confidence intervals for our estimates, that allow us to be confident that we find a true upper-bound.

We empirically validate our findings. First, using random rank-one, and full-rank matrices for the clean and bot distributions respectively, we verify accuracy using Monte Carlo simulations. Second, we examine Twitter (now X) data. Twitter accounts with large follower-ship that were offered for sale on an open market-place are flagged as having >90% bot followers, while accounts for several academic conferences and well-known researchers are flagged at <20%. We verify accuracy on Twitter account populations of arbitrary clean/bot composition.

Invariant-Guided Logical Testing of Open RAN Controllers

Tianchang Yang, Ali Ranjbar, Gang Tan, and Syed Rafiul Hussain, The Pennsylvania State University

Available Media

Open RAN (O-RAN) represents a fundamental shift in mobile network architecture, advancing interoperability and flexibility through open interfaces and software-driven components. While enabling programmability and innovation, this shift also makes the logical correctness of O-RAN components essential for the secure and reliable operation of the network. However, validating O-RAN's semantic correctness remains challenging due to system complexity, implementation diversity, and the absence of explicit correctness oracles. We present InvaRAN, a systematic testing framework for detecting logical flaws in O-RAN implementations using dynamically inferred program invariants as proxies for expected behavior. To reduce false positives and focus on semantically meaningful behaviors, InvaRAN classifies invariants into critical and non-critical categories based on their impact on program logic. Beyond traditional template-based invariant inference approaches that infer only limited semantic relations, InvaRAN captures inter-variable correlations across execution traces to discover more expressive semantic linkage. We evaluate InvaRAN on both platform components and xApps of two production-grade O-RAN controllers. InvaRAN uncovers nine previously unknown issues, including seven logical and two memory vulnerabilities, demonstrating the effectiveness of invariant-guided testing in exposing subtle, specification-silent bugs in O-RAN systems.

Kintsugi: Empowering LLMs to Mitigate Web Vulnerabilities via Runtime Policy Injection

Yihao Peng, Zizhen Zhu, Jiatian Hu, Jiaxu Wang, Hai Wan, and Xibin Zhao, Tsinghua University

Available Media

The "response gap" between vulnerability detection and patching leaves web applications exposed to high-impact exploits such as remote code execution, command injection, and server-side request forgery. Current mitigations are flawed: code fixes often break functionality, while global runtime policies produce excessive false positives. We propose Kintsugi, an automated runtime containment system that provides temporary protection for exploits that manifest as diverging syscall behaviors at the OS level. Kintsugi uses a three-stage pipeline: First, by performing differential syscall analysis on normal and malicious requests, it locates a few vulnerability-related functions. Subsequently, leveraging LLMs, it precisely delineates the boundaries of the specific code snippets causing the malicious behavior within these functions and establishes policy trigger points. Finally, by analyzing the execution profiles from normal requests, it derives a deterministic, least-privilege syscall whitelist to serve as the runtime policy. This policy is enforced at the kernel level via eBPF and cgroups and is dynamically activated only when the request's execution flow enters the protected code snippet. Evaluation on 27 real-world CVEs across PHP, Python, and Java shows that Kintsugi effectively neutralizes diverse exploits and their variants while preserving original application functionality. Kintsugi achieves an average response time of 7.6 minutes. While the enforcement of surgical policies introduces a modest average latency overhead of 9.2% on targeted APIs, the impact on concurrent non-vulnerable traffic remains minimal, with an average throughput (RPS) drop of only 2.21%.

Alias Equals Zone? Large-Scale and Stealthy Takeover of Domain Hosting Service via CNAME-Following Cross-Domain Verification

Ruixuan Li, Tsinghua University; Xingyu Zhao, Zhejiang Gongshang University; Yunyi Zhang and Baojun Liu, Tsinghua University; Jun Shao, Zhejiang Gongshang University and Zhejiang Key Laboratory of Big Data and Future E-Commerce Technology

Available Media

CNAME records define alias relationships between domains and are widely used for service hosting and load balancing. We find that popular domain hosting providers misinterpret CNAME semantics during domain ownership verification. They accept DNS records after CNAME redirection as valid challenge tokens for alias domains, even though these domains do not configure any tokens. Based on this flaw, we propose ALIASLEAP, a novel domain takeover attack that enables hijacking hosting services of alias domains in CNAME chains. ALIASLEAP poses a serious threat in the real world: we identify four email and seven web hosting providers that are vulnerable, affecting over two million domains, including 200K in the Tranco Top 1M domain list. ALIASLEAP is highly stealthy because vulnerable CNAME chains are typically legitimate and long-lived: about half persist for more than 12 months, and up to 19,819 domains have been exposed for over 10 years. We propose mitigation strategies and responsibly disclose ALIASLEAP to 11 affected hosting providers, receiving confirmations from eight of them. We call on the Internet community to revisit the usage practices and capability boundaries of CNAME records.

Securing Retrieval-Augmented Code Generation via Contextual Knowledge Injection: A Case for Embedded IoT Applications

Tong Sun, Jingyi Su, Yi Gao, and Wei Dong, The State Key Laboratory of Blockchain and Data Security, College of Computer Science, Zhejiang University

Available Media

Repository-grounded retrieval-augmented code generation (RACG) is increasingly used in embedded IoT development by retrieving code and documentation from a pinned RTOS/SDK repository (e.g., Zephyr OS). In this setting, security risks are often version-inherited: even without retrieval poisoning, generated applications may invoke benign-looking public APIs that transitively reach vulnerable internal routines in the pinned snapshot, thereby inheriting known CVEs. Existing secure RACG pipelines largely focus on task-level intent and generic vulnerability patterns, which can miss repository- and version-specific exposure. Meanwhile, conventional CVE scanners can flag vulnerable locations but cannot determine whether those vulnerabilities are reachable through the public APIs that the generator commits to during repository-grounded generation.

In this paper, we present IoTRAGuarder, a contextual knowledge injection framework that aligns security hardening with generation-time API selection under repository grounding. IoTRAGuarder (i) recovers auditable reverse call chains from CVE-localized internals to exposing public APIs via static analysis plus an evidence-gated LLM to bridge indirections and macro-driven "call-graph islands", (ii) constructs a version-aware security knowledge base that binds affected version intervals to exposed public APIs with prompt-ready constraints, safer alternatives, or avoidance/upgrade guidance, and (iii) performs dual-layer, API-aligned online retrieval to inject concise, version-matched constraints into the final prompt. We evaluate IoTRAGuarder on 44 real-world Zephyr tasks across four LLMs. Compared to the prior state-of-the-art secure RACG baseline, IoTRAGuarder improves the overall security success rate from 5.11% to 78.41%.

Track 2

System Security 1

Session Chair: Dario Stabili, University of Modena and Reggio Emilia

Grand Ballroom VI

KernelRCA: Facilitating Root Cause Analysis of Memory Corruptions in Linux Kernel with Contextual Causality Chain

Kangzheng Gu, Yifan Zhang, Yuan Zhang, and Min Yang, Fudan University

Available Media

Continuous fuzzing infrastructure has found a large number of bugs. In this case, automatic root cause analysis (RCA) has been proposed to reduce the expensive manual effort to understand the root cause of a bug. However, existing root-cause representations are designed as isolated forms. Analysts still need to manually infer the integrated bug-triggering procedure including calling context and data dependency, which is very difficult for OS kernels due to their complexity.

In this paper, we propose contextual causality chain (CC-chain), a novel root-cause representation to intuitively reflect the integrated bug-triggering procedure of memory corruptions in the Linux kernel. CC-chain shows the bug-contributing instructions to explain corresponding unexpected behaviors that lead to a bug, as well as calling contexts and data dependencies among these instructions to help analysts rapidly understand how a bug happens. To automatically construct the CC-chain, we design a root cause analysis system KernelRCA including selective tracing, contextual information recovery, and chain-style root cause analysis. KernelRCA successfully diagnoses 54 various kinds of real-world memory corruptions in the Linux kernel and performs better than existing crash reports and KASAN reports. A user study shows that KernelRCA's reports significantly facilitate bug understanding and fixing for human analysts.

Discovering, characterizing and exploiting controllable-copy objects for kernel data-only attacks with CopyKat

Jakob Koschel, Andrea Mambretti, Alessandro Sorniotti, Pietro Moretto, and Claudio Migliorelli, IBM Research Europe; Andrea Di Dio and Cristiano Giuffrida, Vrije Universiteit Amsterdam; Anil Kurmus, IBM Research Europe

Available Media

Modern kernel hardening mechanisms considerably reduce the viability of control-flow hijacking attacks, prompting attackers to shift their focus to data-only attacks. These attacks depend on corrupting non-control data to amplify limited memory-corruption primitives into more powerful capabilities such as arbitrary writes. Prior work has identified small sets of such objects in specific circumstances. However, no existing approach discovers and characterizes them at scale. In this paper, we introduce CopyKat, a fully automated pipeline for identifying and characterizing Controllable-Copy Objects (CCOs): heap-allocated kernel structures whose fields can be corrupted to redirect a store or memory-transfer operation involving attacker-controlled data. We propose a multi-stage methodology with progressive filtering of objects. Through lightweight taint-analysis applied to a fuzzing campaign, we uncover candidate CCOs. Then, we confirm such taint by using a precise taint engine, thereby enhancing our precision. Finally, we perform the corruption of the CCO using a concolic execution engine to verify the controllability of the object. We also collect the necessary path and value constraints for exploitation. We find 122 verified CCOs and build 8 end-to-end exploits involving different exploitation patterns, showcasing the correctness of our characterization and the flexibility of the discovered objects.

Static Detection of TOCTOU Bugs Caused by Kernel Races

Gui-Dong Han, Jia-Ju Bai, Qiu-Ji Chen, and Jiqiang Lu, Beihang University

Available Media

The TOCTOU (Time Of Check to Time Of Use) bug is a well-known security issue in kernel code, because it bypasses security checks and leads to unexpected behaviors that can cause serious problems like system crashes and privilege escalation. According to our study on Linux kernel patches, kernel race is the most common root cause of kernel TOCTOU bugs. However, due to the complexity of kernel concurrency logic and non-determinism of thread scheduling, there is still no systematic approach that focuses on detecting TOCTOU bugs caused by kernel races.

In this paper, we design KERAT, the first systematic static approach for detecting TOCTOU bugs caused by kernel races. Indeed, such TOCTOU bugs are introduced by atomicity violations about the check-use operations of specific shared variables. Thus, KERAT performs bug detection by statically mining and checking the atomicity rules about shared variables from kernel code. Specifically, KERAT has two key techniques: (1) an atomicity-rule mining method to effectively identify which lock should protect the check-use operations of which shared variable; and (2) a state-based validation strategy to detect TOCTOU bugs that violate the mined atomicity rules based on state machines encoding of common bug patterns. We have evaluated KERAT on Linux-6.8 and FreeBSD-14.1, and found 351 real bugs. Among these bugs, 287 are identified as harmful, and 65 of them have been confirmed by kernel developers. 10 bugs have received CVE IDs.

SoK: On the Fragility of Memory Error Exploit Mitigations

Adriaan Jacobs, DistriNet, KU Leuven; Mahmoud Ammar, Huawei Research, Germany; Stijn Volckaert, DistriNet, KU Leuven

Available Media

The perennial war in memory has long been shaped by a continuous arms race: defenses are deployed, bypasses emerge, and stronger mitigations follow, only for the cycle to repeat. This pattern persists in part due to the fragility of many defenses, which often fail when assumptions change. Such fragility reflects how security guarantees are assessed, often through ad hoc reasoning tied to specific threat models. In a fast-evolving landscape of polyglot applications and heterogeneous systems, manually re-evaluating these guarantees for every new context is both labor-intensive and error-prone.

This SoK advocates for a more systematic,adversary-aware approach. We introduce a graph-based framework for evaluating the fragility of memory safety defenses by modeling the progression of memory corruption exploits, and where, how, and under what assumptions defenses intervene and may fail. We demonstrate the utility of this model by revisiting flaws in prominent defenses and showing how their shortcomings could have been anticipated. Finally, we release open-source tooling that implements our model and supports systematized and semi-automated fragility testing.

Fence2Pwn: KFENCE-Enabled Kernel Exploitation Bypassing Slab Hardening and Memory Tagging

Ernesto Martínez García, Lukas Maar, Martin Unterguggenberger, and Stefan Mangard, Graz University of Technology

Available Media

While the Linux kernel remains a prime target due to its privileged execution model, exploitation has become substantially more difficult in recent years. Kernel hardening efforts have increasingly focused on the slab allocator, aiming to mitigate entire vulnerability classes (e.g., via memory tagging) or disrupt exploit techniques (e.g., via heap object segregation). However, it remains unclear whether these protections are consistently enforced across all allocation paths.

In this paper, we find that one allocation path is excluded from all state-of-the-art slab defenses: the path used for KFENCE allocations. KFENCE is designed as a low-overhead sampling-based memory-safety error detector for production kernels. It is enabled by default and active in billions of systems including on Android, Red Hat Enterprise Linux, and most Linux distributions. We show that it unintentionally enables a new exploit technique, Fence2Pwn. Concretely, Fence2Pwn leverages several KFENCE-specific behaviors. Three are particularly notable: (i) using a timing side channel, we detect when the kernel falls back to KFENCE allocations; (ii) KFENCE allocations are served by KFENCE-managed caches rather than size- and type-segregated slab caches, enabling bypass of heap segregation; and (iii) KFENCE allocations are untagged, enabling bypass of memory tagging. We evaluate Fence2Pwn by profiling our timing side channel, KFENCE's object slot dynamics as well as different environments and noise floors. Finally, we demonstrate that Fence2Pwn exploits this slab-defense gap: Fence2Pwn uses KFENCE to exploit an existing UAF-based memory-reuse vulnerability, which allows it to overlay security-relevant objects of different types and sizes.

kSFS: Repurposing a Microkernel-like Interface for Fast and Secure In-Kernel Linux File Systems

Dinglan Peng and Pedro Fonseca, Purdue University

Available Media

File systems are widely-used and crucial but notoriously complex and a major source of vulnerabilities in operating systems. Recent works have proposed introducing in-kernel sandboxing techniques to isolate kernel components including file systems. However, a well-defined and secure boundary, where all interactions between untrusted and trusted kernel components should be validated against a strong threat model, is often ignored. This lack of secure boundary particularly applies to Linux file systems, which rely on a large and complex interface and interact with many kernel subsystems such as VFS and block devices. Defining such an interface is a challenging prerequisite of sandboxed kernel file systems.

We address this challenge with kSFS, a framework for in-kernel sandboxed file systems. kSFS repurposes the FUSE protocol, which is a microkernel-like interface originally designed for user-space file systems in Linux, as a secure interface for untrusted sandboxed kernel file systems that has strong isolation guarantees. Furthermore, kSFS generalizes WebAssembly to kernel space as a generic sandboxing mechanism and achieves compatibility with existing user-space file system implementations with minimal porting effort. For instance, porting the NTFS and exFAT implementations from user space with kSFS required modifying fewer than 300 LoC. While achieving better security and reliability than Linux file system implementations, kSFS achieves significantly better performance than their user-space counterparts. For the real-world applications tar and RocksDB, the kSFS NTFS implementation achieves up to 29% and 60× better performance than the user-space baseline, respectively, and only 0% to 52% lower performance than the insecure Linux implementation.

CoKeMon: Configurable Kernel Monitoring by Decoupling Isolation

Clément Thorens and Shweta Shinde, ETH Zurich

Available Media

Monolithic OSes such as Linux are susceptible to security vulnerabilities. This has led to a line of research to detect and prevent runtime kernel exploitation, e.g., with kernel integrity measurement, page table monitoring, event logging. Prior works are either purpose-built for one type of monitoring or make invasive changes to the kernel. This motivates the need for a design that can accommodate various monitors on a need basis. As a first step, we identify five key requirements that such a solution must satisfy: selective permissions, non-privileged management, static isolation boundaries, architecturally-supported atomic switching, and device support. It is challenging to satisfy all of the requirements while preserving security. To address this, our insight is to decouple the isolation enforcement from the monitor management, which allows us to use native hardware techniques with ease. We demonstrate the feasibility of such a design on Arm platform by modifying ~300 LoC in the firmware and ~500 LoC in the Linux kernel. We implement three monitoring use-cases to demonstrate the versatility of COKEMON and report performance overhead ranging from 0% to 13%.

DMGuard: Safeguarding Kernels from Physical-Page Use-After-Free Vulnerabilities

Juhee Kim, Jaeyoung Chung, Dae R. Jeong, and Byoungyoung Lee, Seoul National University

Available Media

Modern kernels depend on the integrity of page tables to enforce advanced security measures. Although these defenses have effectively mitigated various attacks including memory corruption, adversaries have shifted their focus to compromise the page table itself to bypass existing protections. Such threats are exacerbated by the rise of heterogeneous address translation domains including separate CPU, GPU, and IOMMU page tables, which impose heavy demands on synchronization and coherence management. When a virtual address remains mapped to a physical page that has already been freed or reallocated, attackers can exploit this to access arbitrary physical memory. We call this physical-page use-after-free, distinct from traditional heap use-after-free that operates on virtual addresses.

In this paper, we present DMGuard, the first runtime mitigation that comprehensively addresses physical-page use-after-free vulnerabilities across diverse translation domains. DMGuard leverages a lightweight, lockless mechanism to manage a state machine of physical pages to ensure no dangling mappings exist in the page tables. Evaluation of DMGuard on Android devices demonstrates that it effectively blocks all known physical-page use-after-free vulnerabilities with negligible performance overheads, demonstrating the practicality and effectiveness against emerging attacks.

Track 3

Applications of Cryptography 2

Session Chair: Adithya Bhat, Visa Research

Grand Ballroom V

Exploiting PoH Time Semantics in Solana via Re-Anchoring and Forking

Quanbi Feng, Southern University of Science and Technology; Pinshen Xu, Jianyu Niu, and Cong Wang, City University of Hong Kong; Yinqian Zhang, Southern University of Science and Technology

Available Media

Proof of History (PoH) is a core component of Solana that realizes a publicly verifiable notion of logical time via a sequential hash chain. It allows Solana to run a slot-based leader schedule, where a designated leader is expected to propose a block in each slot. Yet, the security implications of PoH-driven logical time remain insufficiently understood. In this paper, we identify a new protocol-valid attack surface in Solana's PoH time semantics: by withholding and later releasing protocol-valid blocks that commit to an earlier PoH-derived logical time than a validator's current local view, a scheduled malicious leader can trigger PoH re-anchoring at honest validators. Building on this primitive, we develop two attacks. First, Time Inflation Attack (TI) extends the malicious leader's effective block-production time budget by multiple slot intervals, enabling it to include more transactions without necessarily invalidating honest blocks. Second, Fork-Assisted Time Inflation Attack (FTI) extends the former by leveraging fork choice to further increase the time budget while suppressing honest leaders' proposals.

We present a formal analysis that characterizes the conditions under which PoH re-anchoring occurs and bounds the adversary's impact under realistic network and stake assumptions. We implemented the attacks in a testbed and reported our findings to the Solana development team. We also analyze on-chain data and report timing and inclusion patterns that are compatible with the incentive channel exploited by TI. Finally, we discuss defenses that harden PoH time semantics and reduce incentives to exploit PoH re-anchoring.

Bridging Bitcoin to Second Layers via BitVM2

Robin Linus Woll, Stanford University and ZeroSync Association; Lukas Aumayr, University of Edinburgh and Common Prefix; Zeta Avarikioti, TU Wien and Common Prefix; Matteo Maffei, TU Wien; Andrea Pelosi, University of Pisa, University of Camerino, and TU Wien; Orfeas Stefanos Thyfronitis Litos, Imperial College London and Common Prefix; Christos Stefo, TU Wien; David Tse, Stanford University and Byzantine Research; Alexei Zamyatin, BOB

Available Media

A holy grail in blockchain infrastructure is a trustless bridge between Bitcoin and its second layers or other chains. We make progress toward this vision by introducing the first light-client-based Bitcoin bridge. At its heart lies BitVM2-CORE, a novel paradigm that enables arbitrary program execution on Bitcoin, combining Turing-complete expressiveness with the security of Bitcoin consensus. BitVM2-BRIDGE advances prior approaches by reducing the trust assumption from an honest majority (t-of-n) to existential honesty (1-of-n) during setup. Liveness is guaranteed with only one rational operator, and any user can act as a challenger, enabling permissionless verification. A production-level implementation of BitVM2 has been developed, and a full challenge verification has been executed on the Bitcoin mainnet.

Just One Bit Vector: Complement-Free Ring Confidential Transactions with Transparent Setup

Hao Gao and Qianhong Wu, School of Cyber Science and Technology, Beihang University; Hangzhou Innovation Institute of Beihang University; Bo Qin, Renmin University of China; Fudong Wu, School of Cyber Science and Technology, Beihang University; Hangzhou Innovation Institute of Beihang University; Zhenyang Ding, School of Cyber Science and Technology, Beihang University

Available Media

Ring Confidential Transactions (RingCTs) rely on their signature schemes to ensure transaction validity while preserving user privacy. In the discrete-logarithm setting, state-of-the art RingCT signatures leverage Bulletproof as their backbone. In this work, we conduct a thorough analysis of existing Bulletproof-based RingCT signatures and identify two pervasive sources of inefficiency: (i) redundant complement advice induced by binary constraints, and (ii) substantial overhead incurred by defenses against the Knowledge-of-Known-Discrete-Logarithm (KKDL) issue. We present new techniques that eliminate the first inefficiency and significantly reduce the second. We introduce the Weighted-Norm Inner Linear Argument (WNILA) as a new compression pivot and leverage it to design zero-knowledge proofs that capture constraints arising in RingCT without introducing complement advice. We then adopt the techniques of Omniring and Lin et al. to unify the K-Weight and K-Link techniques proposed by ZGSX23 and BulletCT. Building on this framework, we derive three RingCT signature schemes: UniBit, UniBit+, and UniBitRange. We implement all three schemes and benchmark them against BulletCT and Omniring⋆ under practical parameter regimes. Our experimental results demonstrate 2.540× improvements in signing and verification speed over Omniring⋆, and 1.6–2.3× signing speedups together with up to 1.65× verification speedup over BulletCT, while achieving comparable signature sizes.

FABS: Fast Attribute-Based Signatures

Liqun Chen, Long Meng, Yalan Wang, Nada El Kassem, Christopher JP Newton, Yangguang Tian, Jodie Knapp, Constantin Cătălin Drăgan, and Daniel Gardham, University of Surrey; Mark Manulis, Universität der Bundeswehr München

Available Media

Attribute-based signatures (ABS) provide fine-grained control over who can generate digital signatures and have many real-world applications. This paper presents a pair of fast ABS schemes: one for Key-Policy ABS (KP-ABS) and another for Signature-Policy ABS (SP-ABS). Both schemes support expressive policies using Monotone Span Programs (MSP), and offer practical features such as large universe, arbitrary attributes, and adaptive security. Most notably, we provide the first implementation of MSP-based ABS schemes and demonstrate that our schemes achieve the best-known asymptotic and concrete performance in this domain. Asymptotically, key generation, signing and verification time scale linearly with the number of attributes; verification requires only two pairing operations. In concrete terms, for 100 attributes, our KP-ABS scheme performs key generation, signing, and verification in 0.16s, 0.10s, and 0.13s, respectively; our SP-ABS scheme achieves times of 0.082s, 0.26s, and 0.21s for the same operations.

The Perils of Flexibility: Uncovering Application-Layer Vulnerabilities in Cosmos SDK Customization Points

Pengxiang Ma and Ningyu He, The Hong Kong Polytechnic University; Zhongchun Cao, ChainFlag; Zihao Li, The Hong Kong Polytechnic University and University of Electronic Science and Technology of China; Xiapu Luo, The Hong Kong Polytechnic University

Available Media

The Cosmos SDK is a premier blockchain development framework that facilitates the construction of application-layer logic through reusable modules and extensible customization interfaces. Currently, the Cosmos ecosystem encompasses over one hundred production blockchains, reaching a peak total market capitalization exceeding $32 billion. While this modular architecture significantly lowers development barriers, it simultaneously expands the application-layer attack surface, as developer-accessible customization points are susceptible to misuse or incorrect composition.

In this work, we present a security analysis of application-layer pitfalls within Cosmos-based blockchains. By analyzing the transaction execution lifecycle and developer-accessible customization points, we formalize four execution invariants essential for correct transaction processing. Guided by these invariants, we identify eight recurring classes of error-prone development patterns in which violations compromise transaction safety, resulting in severe consequences ranging from asset theft to chain-halting failures. To assess the prevalence and impact of these pitfalls, we conducted an ecosystem-wide empirical analysis, uncovering 65 vulnerabilities across various Cosmos-based blockchains. Following responsible disclosure, 26 of these vulnerabilities were acknowledged by the affected project teams, with 19 subsequently receiving public fixes and two resulting in assigned CVE identifiers. These disclosures yielded $51,000 in bug bounty rewards and prevented potential asset theft affecting assets valued in the tens of millions of dollars. Our findings demonstrate that application-layer vulnerabilities are widespread in the Cosmos ecosystem and underscore the critical need for increased rigor in application-layer design, development, and auditing.

Logos: Robust Sharding Blockchain With Fast Processing and Optimal Cross-Shard Overhead

Yizhong Liu, Beihang University and Beijing Advanced Innovation Center for Future Blockchain and Privacy Computing; Boyu Zhao, Yuxuan Hu, Haojun Tan, Feiang Ran, Andi Liu, and Zhuocheng Pan, Beihang University; Yuan Lu, Institute of Software, Chinese Academy of Sciences; Song Bian, Jianwei Liu, and Zhenyu Guan, Beihang University

Available Media

Sharding blockchains improve scalability significantly by partitioning the network into shards. Due to the substantial fraction of cross-shard transactions (CSTXs) related to multiple shards, cross-shard transaction processing (CSTP) is critical to the system security and performance. However, existing CSTP methods suffer from limited robustness caused by invalid CSTXs flooding by malicious nodes and impose high overhead, especially in asynchronous networks.

We present Logos, a robust sharding blockchain with fast CSTP and optimal cross-shard overhead. Logos adopts a novel robust broadcast-transmission-agreement pattern. Each input shard only invokes a new designed broadcast primitive to generate input availability states. After the states are delivered to involved shards by an innovative parallel single-tosingle transmission mechanism, valid CSTXs are committed via an agreement protocol while invalid ones are discarded. Logos is proven to achieve an optimal intra-shard overhead for valid CSTP and lower overhead for invalid CSTP. Besides, Logos achieves reliable transmission with optimal cross-shard overhead. Experiments conducted on 1000 AWS-EC2 nodes across 4 regions demonstrate that Logos realizes 50% latency compared to the baseline (Kronos, NDSS'25) and a peak throughput of 132.8 ktx/sec. Besides, the cross-shard network usage of Logos impressively remains only 1/210 of Kronos. Under malicious flooding attacks, Logos maintains 2.86× the throughput of Kronos, demonstrating strong robustness.

MHOT: Height-Optimized Authenticated Data Structure for Blockchain State Commitment

Sipeng Xie, Qianhong Wu, Minghang Li, and Qiyuan Gao, Beihang University; Bo Qin, Renmin University of China; Qin Wang, Independent

Available Media

State root computation dominates (~78%) blockchain block processing time. Ethereum's canonical authenticated data structure, i.e., Merkle Patricia Trie (MPT), suffers from severe tree-height growth and is vulnerable to Nurgle attacks (S&P'24), where adversaries inflate path depth via hash collisions and degrade system performance at negligible cost. Existing defenses increase node fanout (span) to bound tree height, but higher span inflates proof size exponentially. Prior work mitigates this trade-off using vector commitments, at the cost of trusted setup or expensive verification.

We present MHOT, a height-optimal authenticated data structure for blockchain state commitment that preserves standard hash-based verification without trusted setup. Unlike MPT's fixed-prefix indexing, which couples span and fanout exponentially, MHOT indexes by discriminative bits that actually distinguish keys, achieving adaptive span with linear fanout coupling and provably minimal height. To prevent high fanout from inflating proofs, we introduce hierarchical proofs, a two-layer Merkle construction that reduces per-node proof overhead from O(k) to O(logk).

On Ethereum mainnet workloads, MHOT achieves up to 9× higher write throughput, 4× lower write amplification, and 2× smaller proofs than MPT. Under Nurgle attacks, even when the adversary consumes an entire block's gas budget, MHOT maintains a 0% attack success rate (v.s., 99.97% for MPT). Our results, somewhat surprisingly, show that height optimality (not new crypto primitives!) is the key abstraction for scalable and attack-resilient blockchain state commitment.

Perils of Parallelism: Transaction Fee Mechanisms under Execution Uncertainty

Sarisht Wadhwa, Duke University; Aviv Yaish, Yale University, IC3, Complexity Science Hub Vienna; Fan Zhang, Yale University, IC3; Kartik Nayak, Duke University

Available Media

Modern blockchains increasingly rely on parallel execution to improve throughput. We show several industry and academic transaction fee mechanisms (TFMs) struggle to simultaneously account for execution parallelism while remaining performant and fair. First, if parallelism affects fees, adversarial protocol manipulations that offset possible benefits to throughput by introducing fake transactions become rational: users can insert functionally useless parallel transactions solely to reduce fees, and schedulers can create useless sequential transactions to increase revenue. Execution contingency, a core feature of expressive programming languages, both exacerbates the aforementioned threats and introduces new ones: (1) users may overpay for unused resources, and (2) scheduler revenue is harmed when reserved scheduling slots go unused due to contingency. We introduce a framework for this challenging setting, and prove an impossibility highlighting an inherent tension: both parallelism and contingency involve a trade-off between minimizing risks for users and schedulers, as favoring one comes at the expense of the other. To complete the picture, we introduce fee mechanisms and prove that they achieve the boundaries of this trade-off. Our results provide rigorous foundations for evaluating designs advanced by notable blockchains, such as Sui and Monad.

Track 4

Privacy of ML 1

Session Chair: Sayanton Dibbo, University of Alabama

Grand Ballroom I-IV

Can we estimate privacy vulnerability of individual records? Towards Mitigating Attribute Inference Attacks on ML Models

Ehsanul Kabir and Najrin Sultana, Pennsylvania State University; Ninghui Li, Purdue University; Shagufta Mehnaz, Pennsylvania State University

Available Media

Machine learning (ML) has brought transformative applications across various sectors, including sensitive fields like healthcare, finance, and customer analytics. However, ML models are susceptible to privacy leaks, especially through attribute inference and model inversion attacks, raising concerns for data confidentiality in privacy-critical domains. Existing defenses pursue much broader objectives than specifically preventing privacy leakage from attribute inference attacks, and as a result often fail to provide fine-grained, vulnerability-aware protection without significant utility costs. Motivated by this need, we first investigate record-level vulnerability estimation through NeighVE, an adversary-side tool designed to identify which individual records are more exposed to inference. Insights from NeighVE reveal that the record-level risk of privacy leakage is largely agnostic to model architectures and attack strategies and is instead governed by dataset-level characteristics, particularly the distribution of sensitive attributes in the local neighborhood of each record. Building on this insight, we propose VESL, a subspace-learning–inspired defense that mitigates attribute-inference leakage while keeping utility loss to a bare minimum. As a byproduct of its balancing mechanism, VESL also improves fairness across sensitive attributes and prevents NeighVE from reliably identifying vulnerable records. As a supporting contribution, we introduce AttriVET, an estimator that predicts which individual records are vulnerable with over 90% accuracy across diverse scenarios, enabling risk-aware defense design and auditing.

Sok: Private Transformer-based Model Inference

Yuntian Chen, Northwest University; Xi'an Key Laboratory of Advanced Computing and Software Security; Shaanxi International Joint Research Center for the Battery-Free Internet of Things; Tianpei Lu, The State Key Laboratory of Blockchain and Data Security, Zhejiang University; Hangzhou High-Tech Zone (Binjiang) Institute of Blockchain and Data Security; Zhanyong Tang, Northwest University; Xi'an Key Laboratory of Advanced Computing and Software Security; Shaanxi Key Laboratory of Passive Internet of Things and Neural Computing; Bingsheng Zhang, The State Key Laboratory of Blockchain and Data Security, Zhejiang University; Hangzhou High-Tech Zone (Binjiang) Institute of Blockchain and Data Security; Zhiying Shi, Northwest University; Xi'an Key Laboratory of Advanced Computing and Software Security; Shaanxi Key Laboratory of Passive Internet of Things and Neural Computing; Yuxiang Luan, Northwest University; Xi'an Key Laboratory of Advanced Computing and Software Security; Zhuzhu Wang, Northwest University; Xi'an Key Laboratory of Advanced Computing and Software Security; Shaanxi International Joint Research Center for the Battery-Free Internet of Things

Available Media

The growing demand for privacy-preserving Transformer inference has led to the emergence of numerous protocols designed to protect sensitive data and model parameters. These protocols utilize diverse cryptographic tools under varying assumptions, each presenting unique characteristics and trade-offs between computation, communication, and accuracy. In this paper, we conduct a systematic and in-depth analysis of existing approaches from diverse performance perspectives, identifying their limitations and research gaps. We further evaluate the reproducibility of prior systems and re-benchmark representative solutions under standardized configurations. Our results yield a principled guideline for balancing protocol trade-offs under different deployment settings.

Five Queries Are Enough: Query-Efficient and Surrogate-Free Membership Inference Attacks on RAG via Entailment

Nguyen Linh Bao Nguyen, Wanlun Ma, and Viet Vo, Swinburne University of Technology, Australia; Alsharif Abuadbba, CSIRO, Australia; Minghong Fang, University of Louisville, USA; Jun Zhang and Yang Xiang, Swinburne University of Technology, Australia

Available Media

Retrieval‑Augmented Generation (RAG) has become central to large language model (LLM) deployments, grounding responses in enterprise or proprietary data to reduce hallucinations. However, this design introduces a new privacy risk: model outputs may signal the presence of specific documents in the retrieval corpus, enabling membership inference attacks (MIAs) that leak sensitive information. Existing MIAs are feasible, but they often rely on easily-detected templated queries or require many non-templated yet costly and repetitive queries, limiting practicality. We ask: Can an adversary launch a limited-budget, surrogate-free, stealthy, and defense-agnostic membership inference attack using non-templated queries? We present MEntA (Membership Entailment Attack)—a query‑efficient MIA that leverages natural‑language entailment to maximize information gained per query. By asking low-cost, broad, information‑seeking questions and measuring entailment between model responses and candidate documents, MEntA eliminates the need for costly shadow LLMs and large query budgets. Across NFCorpus, SCIDOCS, and TREC‑COVID, MEntA achieves up to 0.991 AUC with only 5 queries, outperforming prior methods by up to 0.42 AUC under equivalent conditions. It remains effective under state‑of‑the‑art (SOTA) RAG defenses, while current detectors miss MEntA or flag benign queries at high rates. Regarding cost, MEntA reduces total attack cost by up to 65× lower compared to SOTA attacks under the same attack setting. Our findings expose the feasibility of realistic, low‑cost privacy leakage in RAG systems and highlight the urgent need for privacy‑aware retrieval and defense mechanisms.

Imitative Membership Inference Attack

Yuntao Du and Yuetian Chen, Purdue University; Hanshen Xiao, Purdue University & NVIDIA Research; Bruno Ribeiro and Ninghui Li, Purdue University

Available Media

A Membership Inference Attack (MIA) assesses how much a target machine learning model reveals about its training data by determining whether specific query instances were part of the training set. State-of-the-art MIAs rely on training hundreds of shadow models that are independent of the target model, leading to significant computational overhead. In this paper, we introduce Imitative Membership Inference Attack (IMIA), which employs a novel imitative training technique to strategically construct a small number of target-informed imitative models that closely replicate the target model's behavior for inference. Extensive experimental results demonstrate that IMIA substantially outperforms existing MIAs in various attack settings while only requiring less than 5% of the computational cost of state-of-the-art approaches.

Window-based Membership Inference Attacks Against Fine-tuned Large Language Models

Yuetian Chen, Yuntao Du, and Kaiyuan Zhang, Purdue University; Ashish Kundu, Cisco Research; Charles Fleming, Cisco Systems; Bruno Ribeiro and Ninghui Li, Purdue University

Available Media

Most membership inference attacks (MIAs) against Large Language Models (LLMs) rely on global signals, like average loss, to identify training data. This approach, however, dilutes the subtle, localized signals of memorization, reducing attack effectiveness. We challenge this global-averaging paradigm, positing that membership signals are more pronounced within localized contexts. We introduce WBC (Window-Based Comparison), which exploits this insight through a sliding window approach with sign-based aggregation. Our method slides windows of varying sizes across text sequences, with each window casting a binary vote on membership based on loss comparisons between target and reference models. By ensembling votes across geometrically spaced window sizes, we capture memorization patterns from token-level artifacts to phrase-level structures. Extensive experiments across eleven datasets demonstrate that WBC substantially outperforms established baselines, achieving higher AUC scores and 2–3× improvements in detection rates at low false positive thresholds. Our findings reveal that aggregating localized evidence is fundamentally more effective than global averaging, exposing critical privacy vulnerabilities in fine-tuned LLMs.

Membership Inference Attacks on Tokenizers of Large Language Models

Meng Tong, University of Science and Technology of China; Yuntao Du, Purdue University; Kejiang Chen and Weiming Zhang, University of Science and Technology of China; Ninghui Li, Purdue University

Available Media

Membership inference attacks (MIAs) are widely used to assess the privacy risks associated with machine learning models. However, when these attacks are applied to pre-trained large language models (LLMs), they encounter significant challenges, including mislabeled samples, distribution shifts, and discrepancies in model size between experimental and real-world settings. To address these limitations, we introduce tokenizers as a new attack vector for membership inference. Specifically, a tokenizer converts raw text into tokens for LLMs. Unlike full models, tokenizers can be efficiently trained from scratch, thereby avoiding the aforementioned challenges. In addition, the tokenizer's training data is typically representative of the data used to pre-train LLMs. Despite these advantages, the potential of tokenizers as an attack vector remains unexplored. To this end, we present the first study on membership leakage through tokenizers and explore five attack methods to infer dataset membership. Extensive experiments on millions of Internet samples reveal the vulnerabilities in the tokenizers of state-of-the-art LLMs. To mitigate this emerging risk, we further propose an adaptive defense. Our findings highlight tokenizers as an overlooked yet critical privacy threat, underscoring the urgent need for privacy-preserving mechanisms specifically designed for them.

VidLeaks: Membership Inference Attacks Against Text-to-Video Models

Li Wang and Wenyu Chen, Shandong University; Ning Yu, Eyeline Labs; Zheng Li and Shanqing Guo, Shandong University

Available Media

The proliferation of powerful Text-to-Video (T2V) models, trained on massive web-scale datasets, raises urgent concerns about copyright and privacy violations. Membership inference attacks (MIAs) provide a principled tool for auditing such risks, yet existing techniques—designed for static data like images or text—fail to capture the spatio-temporal complexities of video generation. In particular, they overlook the sparsity of memorization signals in keyframes and the instability introduced by stochastic temporal dynamics.

In this paper, we conduct the first systematic study of MIAs against T2V models and introduce a novel framework VidLeaks, which probes sparse-temporal memorization through two complementary signals: 1) Spatial Reconstruction Fidelity (SRF), using a Top-K similarity to amplify spatial memorization signals from sparsely memorized keyframes, and 2) Temporal Generative Stability (TGS), which measures semantic consistency across multiple queries to capture temporal leakage. We evaluate VidLeaks under three progressively restrictive black-box settings—supervised, reference-based, and query-only. Experiments on three representative T2V models reveal severe vulnerabilities: VidLeaks achieves AUC of 82.92% on AnimateDiff and 97.01% on InstructVideo even in the strict query-only setting, posing a realistic and exploitable privacy risk. Our work provides the first concrete evidence that T2V models leak substantial membership information through both sparse and temporal memorization, establishing a foundation for auditing video generation systems and motivating the development of new defenses. Code is available at: https://zenodo.org/records/17972831.

Inference Attacks Against Graph Generative Diffusion Models

Xiuling Wang and Xin Huang, Hong Kong Baptist University; Guibo Luo, Peking University; Jianliang Xu, Hong Kong Baptist University

Available Media

Graph generative diffusion models have recently emerged as a powerful paradigm for generating complex graph structures, effectively capturing intricate dependencies and relationships within graph data. However, the privacy risks associated with these models remain largely unexplored. In this paper, we investigate information leakage in such models through three types of black-box inference attacks. First, we design a graph reconstruction attack, which can reconstruct graphs structurally similar to those training graphs from the generated graphs. Second, we propose a property inference attack to infer the properties of the training graphs, such as the average graph density and the distribution of densities, from the generated graphs. Third, we develop two membership inference attacks to determine whether a given graph is present in the training set. Extensive experiments on three different types of graph generative diffusion models and six real-world graphs demonstrate the effectiveness of these attacks, significantly outperforming the baseline approaches. Finally, we propose two defense mechanisms that mitigate these inference attacks and achieve a better trade-off between defense strength and target model utility than existing methods. Our code is available at https://zenodo.org/records/17946102.

Track 5

Enigma Track: Security in the Wild

Session Chair: Joseph Calandrino, Carnegie Mellon University

Dover Room

Verifiable Elections—Now Is the Time

Josh Benaloh, Microsoft Research

Available Media

We have a crisis of confidence in elections today in the U.S. and in many parts of the world. Voters are asked to "trust the process" without being given direct evidence that their votes have been counted correctly. We can do better!

Technology has existed for over four decades that allows voters to independently verify that their votes have been accurately counted – without having to trust any election software, hardware, or personnel. This technology is now mature, practical, easy to use, and has been deployed in public elections.

This talk will describe what it means for an election to be publicly verifiable, provide a high-level view of how these systems work, and reflect on lessons learned from real-world deployments. It will also discuss how individuals can help drive the adoption of verifiable election technology in their own jurisdictions.

Josh Benaloh is the Senior Principal Cryptographer at Microsoft Research and an Affiliate Professor at the University of Washington's Paul G. Allen School of Computer Science & Engineering. He is an author of the 2018 National Academy of Sciences report "Securing the Vote: Protecting American Democracy" and has testified before Congress on verifiable election technologies. His work has been featured in publications including The New Yorker and WIRED. Dr. Benaloh holds an S.B. degree from the Massachusetts Institute of Technology and M.S., M. Phil., and Ph.D. degrees from Yale University. He served seventeen years on the Board of Directors of the International Association for Cryptologic Research and currently serves on the Coordinating Committee of the Election Verification Network which he chaired from 2020-2024. Outside of professional activities, Dr. Benaloh served eight years on and chaired the Citizen Oversight Panel for Sound Transit which is investing $2 billion annually on expanding the public transit infrastructure for the Seattle region. He has also authored numerous puzzles for competitive puzzle-solving events.

If They Can Touch It, They Own It: The Stories We Tell Ourselves About Medical Device Cybersecurity

Michael Rushanan, Armatyr

Available Media

Medical device security differs from traditional security in one important way: when it fails, patient safety can be at risk. Regulators around the world have responded with policies, requirements, and frameworks intended to drive secure-by-design development. However, regulatory frameworks rely heavily on manufacturer risk assessments. As a result, the quality of the underlying security reasoning becomes critical.

This talk presents real-world case studies of security decisions shaped by familiar justifications, including "competitors are doing less," "previous products shipped without incident," and "physical access makes security irrelevant." These decisions result in flawed risk management, weak cryptographic designs, and avoidable vulnerabilities. To illustrate how these decisions emerge in practice, each case study describes the technical issue, stakeholder dynamics, and outcome. It also discusses penetration-testing findings, regulatory feedback, and practical remediation approaches.

Dr. Michael Rushanan is a medical device cybersecurity researcher, consultant, and educator with more than two decades of experience in security engineering, risk management, and applied cryptography. He is the Co-Founder of Armatyr and previously spent 11 years as Chief Scientist at Harbor Labs, helping manufacturers build and evaluate the security of medical devices. Dr. Rushanan is also a Fellow at Harbor Experts and a Lecturer in Computer Science at Johns Hopkins University. He teaches Medical Device Cybersecurity (MDC), a course he designed to address the need for applied education in the field. Dr. Rushanan also directs federal cybersecurity efforts for the U.S. FDA, HHS, and DOJ, and has provided expert testimony on the security of health and medical systems.

When AI Stops Answering and Starts Acting: Field Notes from Red-Teaming the Agentic Era

Sangyoon (David) Yu, AIM Intelligence

Every frontier model we have been given access to, hardened ones included, was jailbroken. That was tolerable when the worst outcome was a bad answer. It no longer is: AI agents act, and we now see sessions where every individual step passes policy while the session as a whole is the attack. Drawing on daily red-teaming of frontier models and Korea's national AI red-team programs, this talk walks through what the field actually looks like: the lethal trifecta showing up in production agents, MCP as a shared supply-chain attack surface, attacks extending past text into pixels, sound, and robots, and why the economics structurally favor the attacker. On defense, I cover why single-wall guardrails provably cannot hold and what works instead: per-action zero trust in a defense-in-depth lattice. I close with evidence that safety does not transfer across languages and countries, and a replicable public-private model for independent adversarial testing.

Sangyoon (David) Yu is CEO and co-founder of AIM Intelligence, a Korean AI-security company that both attacks and defends AI systems: automated red-teaming that finds vulnerabilities in frontier models and agents, and real-time guardrails that control them in production. His team has published 17 papers at venues including ACL, ICML, ICLR, NeurIPS, CVPR, and Nature, partners with OpenAI on guardrails, and co-runs Korea's national AI red-team programs, working with 15+ enterprises and 7+ government agencies across finance, manufacturing, telecom, and the public sector.

2:30 pm–3:00 pm

Coffee and Tea Break

Grand Foyer North and West

3:00 pm–4:30 pm

Track 1

Privacy Anonymity 1

Session Chair: Wei Dong, Nanyang Technological University

Grand Ballroom VII-X

Distributed Synthesis of Differentially Private Tabular Datasets

Yucheng Fu, University of Virginia; Tianyao Gu and Elaine Shi, Carnegie Mellon University; Tianhao Wang, University of Virginia

Available Media

Differentially private synthetic data generation has emerged as a powerful tool for sharing data while protecting individuals' privacy. However, when the attributes of sensitive data are distributed across multiple entities such as hospitals, companies, or government agencies, accurately generating synthetic data becomes challenging. In particular, it is difficult to capture informative statistical correlations and use them to guide data synthesis without gathering the entire private dataset. In response to this challenge, we propose a secure multi-party computation protocol for differentially private tabular data synthesis in the distributed setting. Our protocol contains two new primitives. The first is a protocol that exploits distributed point functions to efficiently estimate two-way marginals (pairwise joint distributions of attributes) across vertically distributed data. The second is a protocol for generating noise via batched lookups in the cumulative distribution function table. As a concrete demonstration, we build a distributed version of AIM, a state-of-the-art DP data-synthesis algorithm. Our implementation achieves the same utility as its centralized version while reducing end-to-end runtime by orders of magnitude compared with prior work. For example, we can synthesize the "Adult" dataset in 24 minutes in a real-world WAN setting, whereas the existing protocol is estimated to take 57 days.

The Adverse Effects of Omitting Records in Differential Privacy: How Sampling and Suppression Degrade the Privacy–Utility Tradeoff

Àlex Miranda-Pascual, Kalrsruhe Institute of Technology and Universitat Politècnica de Catalunya; Javier Parra-Arnau, Universitat Politècnica de Catalunya; Thorsten Strufe, Karlsruhe Institute of Technology

Available Media

Sampling is renowned for its privacy amplification in differential privacy (DP), and is often assumed to improve the utility of a DP mechanism by allowing a noise reduction. In this paper, we further show that this last assumption is flawed: When measuring utility at equal privacy levels, sampling as preprocessing consistently yields penalties due to utility loss from omitting records over all canonical DP mechanisms—Laplace, Gaussian, exponential, and report noisy max— , as well as recent applications of sampling, such as clustering.

Extending this analysis, we investigate suppression as a generalized method of choosing, or omitting, records. Developing a theoretical analysis of this technique, we derive privacy bounds for arbitrary suppression strategies under unbounded approximate DP. We find that our tested suppression strategy also fails to improve the privacy–utility tradeoff. Surprisingly, uniform sampling emerges as one of the best suppression methods—despite its still degrading effect. Our results call into question common preprocessing assumptions in DP practice.

High-Accuracy, Poisoning-Resilient Frequency Estimation in the Shuffle Model

Shaoqiang Wu, CS&CCS, DISSec, AAIS, Nankai University; Jingyu Jia, CS&CCS, DISSec, AAIS, Nankai University; Automotive New Technology Research Institute of BYD; Yikuan Zhu and Xinhao Li, CS&CCS, DISSec, AAIS, Nankai University; Changyu Dong, Guangzhou University; Zheli Liu, CS&CCS, DISSec, AAIS, Nankai University

Available Media

We study frequency estimation in the shuffle model of differential privacy under poisoning attacks, where corrupted users may deviate from the local randomizer to inject crafted in-domain messages. Existing shuffle-model protocols face a core tension: achieving low estimation error relies on flexible multi-message noise generation, which can amplify poisoning influence once messages are anonymized by shuffling.

To address this tension, we propose a symmetric binomial-sum noise distribution (i.e., Bin(n/2, p)+Bin(n/2,1− p)), which preserves high accuracy while limiting the impact of crafted in-domain messages. We realize this distribution via preprocessing-guided noise generation, which routes a balanced collection of mode flags through the shuffler so that each user receives a randomly assigned mode flag that fixes their noise-sampling behavior prior to shuffling. For binary estimation, our protocol requires a single Bernoulli trial per user and at most 2 messages per user(1.5 on average),while bounding the worst-case poisoning influence of a single corrupted user by O(1/n). We extend the protocol to histograms, including large domains via hashing, and provide formal privacy, accuracy, and robustness guarantees. Experiments on real datasets show that our protocols remain resilient under poisoning and reduce MAE by up to nearly 2× over the strongest baseline at comparable per-user communication on small domains, and match it on large domains.

Residual-PAC Privacy: Automatic Privacy Control Beyond the Gaussian Barrier

Tao Zhang and Yevgeniy Vorobeychik, Washington University in St. Louis

Available Media

The Probably Approximately Correct (PAC) Privacy framework [xiao2023pac] provides a powerful instance-based methodology to preserve privacy in complex data-driven systems. Existing PAC Privacy algorithms (we call them Auto-PAC) rely on a Gaussian mutual information upper bound. However, we show that the upper bound obtained by Auto-PAC is tight if and only if under the data distribution, the unperturbed output is Gaussian and the noise is independent Gaussian. We propose two approaches for addressing this issue. First, we introduce two tractable post‐processing methods for Auto-PAC, based on Donsker–Varadhan representation and sliced Wasserstein distances. However, the result still leaves "wasted" privacy budget. To address this issue more fundamentally, we introduce Residual-PAC (R-PAC) Privacy, an f-divergence-based measure to quantify privacy that remains after adversarial inference. To implement R-PAC Privacy in practice, we propose a Stackelberg Residual-PAC (SR-PAC) automatic privatization algorithm, a game-theoretic framework that selects optimal noise distributions through convex bilevel optimization. Our approach achieves efficient privacy budget utilization for arbitrary data distributions and naturally composes when multiple mechanisms access the dataset. Our experiments demonstrate that SR-PAC obtains consistently a better privacy-utility tradeoff than both PAC and differential privacy baselines.

Fully Oblivious Differential Privacy for Frequency Estimation in the Augmented Shuffle Model with Trusted Processors

Takao Murakami, ISM/ROIS/AIST/RIKEN AIP; Yuichi Sei, UEC; Reo Eriguchi, AIST

Available Media

In the shuffle model of DP (Differential Privacy), a shuffler randomly permutes users' data to achieve high accuracy and privacy. Recent studies show that most existing shuffle protocols are vulnerable to collusion attacks by the data collector and users. They address this issue by introducing the augmented shuffle model that incorporates random sampling and dummy data addition into the shuffler. However, it remains open how to ensure the shuffler follows the protocol and does not collude with the data collector in this model.

We address this trust issue by thoroughly exploring the augmented shuffle model with TEEs (Trusted Execution Environments). We first introduce a new privacy notion, FODP (Fully Oblivious DP), which strengthens DP to prevent various TEE side-channel attacks based on external/internal memory access patterns and control flows. We propose a general framework for FODP algorithms based on memory-size obfuscation and three concrete algorithms within it. We also improve the efficiency of our algorithms by using the count-min sketch and optimizing the number of hashes. We evaluate our algorithms on Intel SGX and demonstrate their effectiveness through comparisons with nine baselines.

Sequential Auditing for f-Differential Privacy

Tim Kutta, Aarhus University; Martin Dunsche, Ruhr University Bochum; Yu Wei and Vassilis Zikas, Georgia Institute of Technology

Available Media

We present new auditors to assess Differential Privacy (DP) of an algorithm based on output samples. Such empirical auditors are common to check for algorithmic correctness and implementation bugs. Most existing auditors are batch-based or targeted toward the traditional notion of (ε,δ)-DP; typically both. In this work, we shift the focus to the highly expressive privacy concept of f-DP, in which the entire privacy behavior is captured by a single tradeoff curve. Our auditors detect violations across the full privacy spectrum with statistical significance guarantees, which are supported by theory and simulations. Most importantly, and in contrast to prior work, our auditors do not require a user-specified sample size as an input. Rather, they adaptively determine a near-optimal number of samples needed to reach a decision, thereby avoiding the excessively large sample sizes common in many auditing studies. This reduction in sampling cost becomes especially beneficial for expensive training procedures such as DP-SGD. Our method supports both whitebox and blackbox settings and can also be executed in one-run frameworks.

Interpolation-Based Optimization for Enforcing lp-Norm Metric Differential Privacy in Continuous and Fine-Grained Domains

Chenxi Qiu, University of North Texas

Available Media

Metric Differential Privacy (mDP) generalizes Local Differential Privacy (LDP) by adapting privacy guarantees based on pairwise distances, enabling context-aware protection and improved utility. While existing optimization-based methods reduce utility loss effectively in coarse-grained domains, optimizing mDP in fine-grained or continuous settings remains challenging due to the computational cost of constructing dense perturbation matrices and satisfying pointwise constraints.

In this paper, we propose an interpolation-based framework for optimizing ℓp-norm mDP in such domains. Our approach optimizes perturbation distributions at a sparse set of anchor points and interpolates distributions at non-anchor locations via log-convex combinations, which provably preserve mDP. To address privacy violations caused by naive interpolation in high-dimensional spaces, we decompose the interpolation process into a sequence of one-dimensional steps and derive a corrected formulation that enforces ℓp-norm mDP by design. We further explore joint optimization over perturbation distributions and privacy budget allocation across dimensions. Experiments on real-world location datasets demonstrate that our method offers rigorous privacy guarantees and competitive utility in fine-grained domains, outperforming baseline mechanisms.

Vεrity: Verifiable Local Differential Privacy

James Bell-Clark, Adrià Gascón, Baiyu Li, and Mariana Raykova, Google; Amrita Roy Chowdhury, University of Michigan

Available Media

Local differential privacy (LDP) enables individuals to report sensitive data while preserving privacy. Unfortunately, LDP mechanisms are vulnerable to poisoning attacks, where adversaries controlling a fraction of the reporting users can significantly distort the aggregate output–much more so than in a non-private solution where the inputs are reported directly. In this paper, we present two novel solutions that prevent poisoning attacks under LDP while preserving its privacy guarantees.
Our first solution, Vεrity-Auth, addresses scenarios where the users report inputs with a ground truth available to a third party. The second solution, Vεrity, tackles the more challenging case in which the users locally generate their input and there is no ground truth which can be used to bootstrap verifiable randomness generation.

Track 2

Security of ML 2

Session Chair: Ziqi Zhang, University of Illinois Urbana–Champaign

Grand Ballroom VI

Measuring Real-World Prompt Injection Attacks in LLM-based Resume Screening

Mohan Zhang, The University of North Carolina at Chapel Hill; Yuqi Jia, Duke University; Zhen Tan, Arizona State University; Steven Jiang, hireEZ; Neil Zhenqiang Gong, Duke University; Tianlong Chen, The University of North Carolina at Chapel Hill; Dawn Song, UC Berkeley

Available Media

LLMs are vulnerable to prompt injection attacks. However, this vulnerability has been primarily demonstrated conceptually in academic studies or through a few anecdotal case studies. Its prevalence and impact in real-world LLM-based applications are largely unexplored. In this work, we present the first systematic study of prompt-injection attacks in a widely used application: LLM-based resume screening. Our analysis is based on approximately 200K real-world resumes collected over multiple years by AnonyCom (The company's name is intentionally anonymized for double-blind review purposes). We first design tailored methods to detect prompt injection in resumes. Manual verification on a small-scale dataset demonstrates that our detectors achieve high precision and outperform state-of-the-art general-purpose detectors. We then apply our detector to the full resume dataset and conduct a comprehensive measurement study of real-world prompt injection attacks. Our analysis reveals several intriguing findings: approximately 1% of resumes contain hidden prompt injections; the prevalence of such injected resumes has increased noticeably over the past one to two years; and more than 90% of injected prompts do not use explicit instructions. These results provide the first evidence of large-scale prompt injection in real-world LLM-based applications and lay the groundwork for future studies to understand and mitigate such attacks.

Cordyceps: Covert Control Attacks on LLMs via Data Poisoning

Zedian Shao, Georgia Institute of Technology; Charles Fleming, Cisco Systems; Teodora Baluta, Georgia Institute of Technology

Available Media

Large language models (LLMs) are often fine-tuned on uncurated text datasets that adversaries can poison. Existing poisoning attacks primarily rely on fixed trigger phrases that defenses such as outlier detection, clean-data regularization, or online monitoring can neutralize. In this paper, we propose a data poisoning method that teaches an LLM an information hiding scheme reliably and stealthily through semantic associations between shared knowledge such as facts or concepts and attacker-chosen phrases. The induced hiding scheme can encode and decode arbitrary malicious instructions, thus revealing a new and subtle poisoning-induced vulnerability: covert control attacks.

We precisely characterize covert control attacks and evaluate them across 5 LLMs, 3 backdoor defenses, and 4 prompt injection defenses. With a small poisoned fraction, covert control attacks outperform heuristic-based prompt injection attacks in average attack success rate by about 40% relative to clean fine-tuned models. They also circumvent defenses based on detection and fine-tuning, maintaining up to 93% attack success rate after backdoor defenses and up to 98% after prompt injection defenses. Our code and data are available at https://github.com/Sadcardation/cordyceps.

The Attacker Moves Second: Stronger Adaptive Attacks Bypass Defenses Against LLM Jailbreaks and Prompt Injections

Milad Nasr and Nicholas Carlini, Anthropic; Chawin Sitawarin, Google DeepMind; Sander V. Schulhoff, HackAPrompt, MATS; Jamie Hayes, Google DeepMind; Michael Ilie, HackAPrompt; Juliette Pluto and Shuang Song, Google DeepMind; Harsh Chaudhari, Northeastern University; Ilia Shumailov, AI Sequrity Company; Abhradeep Guha Thakurta, Google DeepMind; Kai Yuanqing Xiao, OpenAI; Andreas Terzis, Anthropic; Florian Tramèr, ETH Zurich

Available Media

How should we evaluate the robustness of language model defenses? Current defenses against jailbreaks and prompt injections (which aim to prevent an attacker from eliciting harmful knowledge or remotely triggering malicious actions, respectively) are typically evaluated either against a static set of harmful attack strings, or against computationally weak optimization methods that were not designed with the defense in mind. We argue that this evaluation process is flawed.

Instead, we should evaluate defenses against adaptive attackers who explicitly modify their attack strategy to counter a defense's design while spending considerable resources to optimize their objective. By systematically tuning and scaling general optimization techniques—gradient descent, reinforcement learning, random search, and human-guided exploration—we bypass 12 recent defenses (based on a diverse set of techniques) with attack success rate above 90% (under adaptive attacks) for most; importantly, the majority of defenses originally reported near-zero attack success rates. We believe that future defense work must consider stronger attacks, such as the ones we describe, in order to make reliable and convincing claims of robustness.

DualSentinel: A Lightweight Framework for Detecting Targeted Attacks in Black-box LLM via Dual Entropy Lull Pattern

Xiaoyi Pang, The Hong Kong University of Science and Technology; Xuanyi Hao, The State Key Laboratory of Blockchain and Data Security, Zhejiang University; Hangzhou High-Tech Zone (Binjiang) Institute of Blockchain and Data Security; Pengyu Liu, The Hong Kong University of Science and Technology; Qi Luo, Shandong University; Song Guo, The Hong Kong University of Science and Technology; Zhibo Wang, The State Key Laboratory of Blockchain and Data Security, Zhejiang University; Hangzhou High-Tech Zone (Binjiang) Institute of Blockchain and Data Security

Available Media

Recent intelligent systems integrate powerful Large Language Models (LLMs) through APIs, but their trustworthiness may be critically undermined by sequence-coercing targeted attacks that force the model to emit an attacker-chosen payload, including malicious URLs, system commands, and specific misinformation. Existing defensive approaches for such threats typically rely on high access rights, impose prohibitive costs, and hinder normal inference, rendering them impractical for real-world scenarios. To solve these limitations, we introduce DualSentinel, a lightweight defense framework that can accurately and promptly detect the activation of targeted attacks alongside the LLM generation process. We first identify a characteristic of compromised LLMs, termed Entropy Lull: when a sequence-coercing targeted attack successfully hijacks the generation process, the LLM exhibits a distinct period of abnormally low and stable token probability entropy, indicating it is following a fixed path rather than making creative choices. DualSentinel leverages this pattern by developing an innovative dual-check approach. It first employs a magnitude and trend-aware monitoring method to proactively and sensitively flag an entropy lull pattern at runtime. Upon such flagging, it triggers a lightweight yet powerful secondary verification based on task-flipping. An attack is confirmed only if the entropy lull pattern persists across both the original and the flipped task, proving that the LLM's output is coercively controlled. Extensive evaluations show that DualSentinel is both highly effective (superior detection accuracy with the lowest false positives) and remarkably efficient (negligible additional cost), offering a truly practical path toward securing deployed LLMs.

Context Contamination in LLM Analysis of Network Security Logs: Poison with Passive Prompt Injection and Mitigation Evaluation

Rabimba Karanjai, PayPal / University of Houston; Yang Lu, University of Houston; Hemanth Hegadehalli Madhavarao, PayPal; Lei Xu, Kent State University; Weidong Shi, University of Houston

Available Media

Large Language Models are increasingly deployed in Security Operations Centers for log analysis tasks including summarization, alert triage, and threat investigation. These systems ingest logs from external-facing services and process network logs as natural language contexts to generate security insights. We demonstrate that this architectural pattern introduces a critical vulnerability: adversaries can embed prompt injection payloads in log-generating fields that persist in storage and are executed when analysts query the LLM, achieving what we term passive prompt injection. We present LogInject, a systematic framework for evaluating these threats. Using LogInject-1.0, a benchmark of 12,847 log entries including 2,569 adversarial samples, we evaluate three production LLMs across four attack objectives: activity concealment, false positive generation, information exfiltration, and output hijacking. Our findings reveal an up to 88.2% attack success rate (83.4% average across models) under the baseline conditions. We introduce Context Stitching, a novel technique that fragments payloads across multiple log entries to evade stateless filters while exploiting LLM long-context reasoning, achieving a 76.4% success rate. As mitigation, we evaluate layered defenses by combining input filtering, prompt hardening, and output validation, demonstrating a 90.4% attack reduction, although 8.4% residual vulnerability persists. Our results establish that LLM-based log analysis creates an inherent confused deputy vulnerability where untrusted data and trusted instructions compete indistinguishably for model attention, requiring defense in-depth architectures and continued human oversight for security-critical decisions.

MUZZLE: Adaptive Agentic Red-Teaming of Web Agents Against Indirect Prompt Injection Attacks

Georgios Syros and Evan Rose, Northeastern University; Brian Grinstead and Christoph Kerschbaumer, Mozilla Corporation; William Robertson, Cristina Nita-Rotaru, and Alina Oprea, Northeastern University

Available Media

Large language model (LLM) based web agents are increasingly deployed to automate complex online tasks by directly interacting with web sites and performing actions on users' behalf. While these agents offer powerful capabilities, their design exposes them to indirect prompt injection attacks embedded in untrusted web content, enabling adversaries to hijack agent behavior and violate user intent. Despite growing awareness of this threat, existing evaluations rely on fixed attack templates, manually selected injection surfaces, or narrowly scoped scenarios, limiting their ability to capture realistic, adaptive attacks encountered in practice.

We present MUZZLE, an automated agentic framework for evaluating the security of web agents against indirect prompt injection attacks. MUZZLE utilizes the agent's trajectories to automatically identify high-salience injection surfaces, and adaptively generate context-aware malicious instructions that target violations of confidentiality, integrity, and availability. Unlike prior approaches, MUZZLE adapts its attack strategy based on the agent's observed execution trajectory and iteratively refines attacks using feedback from failed executions.

We evaluate MUZZLE across diverse web applications, user tasks, and agent configurations, demonstrating its ability to automatically and adaptively assess the security of web agents with minimal human intervention. Our results show that MUZZLE effectively discovers 44 new attacks on 4 web applications with 10 adversarial objectives that violate confidentiality, availability, or privacy properties across different LLMs and agent scaffolds. MUZZLE also identifies novel attack strategies, including 3 cross-application prompt injection attacks and an agent-tailored phishing scenario.

AttriGuard: Defeating Indirect Prompt Injection in LLM Agents via Causal Attribution of Tool Invocations

Yu He and Haozhe Zhu, Zhejiang University; Yiming Li, Nanyang Technological University; Shuo Shao, Zhejiang University; Hongwei Yao, City University of Hong Kong; Zhihao Liu and Zhan Qin, Zhejiang University

Available Media

LLM agents are highly vulnerable to Indirect Prompt Injection (IPI), where adversaries embed malicious directives in untrusted tool outputs to hijack execution. Most existing defenses treat IPI as an input-level semantic discrimination problem, which often fails to generalize to unseen payloads. We propose a new paradigm, action-level causal attribution, which secures agents by asking why a particular tool call is produced. The central goal is to distinguish tool calls supported by the user's intent from those causally driven by untrusted observations. We instantiate this paradigm with AttriGuard, a runtime defense based on parallel counterfactual tests. For each proposed tool call, AttriGuard verifies its necessity by re-executing the agent under a control-attenuated view of external observations. Technically, AttriGuard combines teacher-forced shadow replay to prevent attribution confounding, hierarchical control attenuation to suppress diverse control channels while preserving task-relevant information, and a fuzzy survival criterion that is robust to LLM stochasticity. Across four LLMs and two agent benchmarks, AttriGuard achieves 0% ASR under static attacks with negligible utility loss and moderate overhead. Importantly, it remains resilient under adaptive optimization-based attacks in settings where leading defenses degrade significantly.

Overcoming the Retrieval Barrier: Indirect Prompt Injection in the Wild for LLM Systems

Hongyan Chang, Ergute Bao, Xinjian Luo, and Ting Yu, Mohamed bin Zayed University of Artificial Intelligence

Available Media

Large language models (LLMs) increasingly rely on retrieving information from external corpora, creating a new attack surface: indirect prompt injection (IPI). Previous studies have highlighted this risk but often avoid the hardest step: ensuring that malicious content is actually retrieved. In practice, unoptimized IPI is rarely retrieved under natural queries, which leaves its real-world impact unclear.

We address this challenge by decomposing the malicious content into a trigger fragment that guarantees retrieval and an attack fragment that encodes arbitrary attack objectives. Based on this idea, we design an efficient and effective black-box attack algorithm that constructs a compact trigger fragment to guarantee retrieval for any attack fragement. Our attack requires only API access to embedding models, is cost-efficient (as little as $0.21 per target user query on OpenAI's embedding models), and achieves near-100% retrieval across 11 benchmarks and 8 embedding models (including both open-source models and proprietary services).

Based on this attack, we present the first end-to-end IPI exploits under natural queries and realistic external corpora, spanning both RAG and agentic systems with diverse attack objectives. These results establish IPI as a practical and severe threat: when a user issued a natural query to summarize emails on frequently asked topics, a single poisoned email was sufficient to coerce GPT-4o into exfiltrating SSH keys with over 80% success in a multi-agent workflow. We further evaluate several defenses and find that they are insufficient to prevent the retrieval of malicious text, highlighting retrieval as a critical open vulnerability.

Track 3

Human Aspects 1

Session Chair: Matthias Fassl, George Washington University

Grand Ballroom V

Love, Lies, and Language Models: Investigating AI's Role in Romance-Baiting Scams

Gilad Gressel and Rahul Pankajakshan, Center for Cybersecurity Systems & Networks, Amrita Vishwa Vidyapeetham, Amritapuri; Shir Rozenfeld, Ben Gurion University of the Negev; Ling Li, Ca' Foscari University of Venice; Ivan Franceschini, University of Melbourne; Krishnashree Achuthan, Center for Cybersecurity Systems & Networks, Amrita Vishwa Vidyapeetham, Amritapuri; Yisroel Mirsky, Ben Gurion University of the Negev

Available Media

Romance-baiting scams have become a major source of financial and emotional harm worldwide. These operations are run by organized crime syndicates that traffic thousands of people into forced labor, requiring them to build emotional intimacy with victims over weeks of text conversations before pressuring them into fraudulent cryptocurrency investments. Because the scams are inherently text-based, they raise urgent questions about the role of Large Language Models (LLMs) in both current and future automation.

We investigate this intersection by interviewing 145 insiders and 5 scam victims, performing a blinded long-term conversation study comparing LLM scam agents to human operators, and executing an evaluation of commercial safety filters. Our findings show that LLMs are already widely deployed within scam organizations, with 87% of scam labor consisting of systematized conversational tasks readily susceptible to automation. In a week-long study, an LLM agent not only elicited greater trust from study participants (p=0.007) but also achieved higher compliance with requests than human operators (46% vs. 18% for humans). Meanwhile, popular safety filters detected 0.0% of romance baiting dialogues. Together, these results suggest that romance-baiting scams may be amenable to full-scale LLM automation, while existing defenses remain inadequate to prevent their expansion.

You Know Why, but Still Rely: The Impact of Explainable AI on Trust, Task Load, and Performance in Cybersecurity Decision-Making

Neele Roch, Hannah Sievers, Noé Zufferey, and Verena Zimmermann, ETH Zurich

Available Media

With the increasing digitisation of institutions, the demand for effective cybersecurity measures is rising rapidly. Simultaneously, the complexity and volume of cybersecurity tasks are outpacing the capacity of available practitioners. Leveraging AI to augment human cybersecurity expertise has the potential to reduce complexities and cognitive overload. Transparent and human-understandable insights into AI decisions are not only demanded by governance authorities, such as the EU, but also by practitioners themselves when collaborating with AI in high-risk contexts. We report on a between-subjects study (N = 139) that investigated the effects of explainable AI (XAI) explanations on trust, usability, perceived task load, and collaborative task performance among users with cybersecurity domain knowledge in the context of malicious domain blocking. The provision of explanations in this context did not foster trust; in fact, users with domain knowledge reported lower trust after interaction with XAI. Qualitative results suggest that they apply their own decision-making criteria, and that exposing AI decision boundaries may introduce ambiguity and foster mistrust. Although the inclusion of XAI did not increase perceived task load, it also failed to improve performance. These findings raise important questions about the effectiveness of current XAI approaches in knowledge-centric, decision-making settings and underscore the need for more context-sensitive, user-aligned explanation strategies in cybersecurity.

"Abuse Risks are Often Inherent to Product Features": Exploring AI Vendors' Bug Bounty and Responsible Disclosure Policies

Yangheran Piao, Jingjie Li, and Daniel W. Woods, University of Edinburgh

Available Media

As vendors adopt AI technologies, security researchers are working to uncover and fix related vulnerabilities, which is important given AI systems handle sensitive data and critical functions. This process relies on vendors receiving and rewarding AI vulnerability reports. To assess current practices, we analyzed the vulnerability disclosure policies of 264 AI vendors. We employed a mixed-methods approach, combining snapshot and longitudinal qualitative analysis, as well as comparing alignment with 320 AI incidents and 260 academic articles. Our analysis reveals that 36% of AI vendors have no established policy, and only 18% mention AI risks. Data access, authorization, and model extraction vulnerabilities are most consistently declared in-scope. Jailbreaking and hallucination are most commonly declared out-of-scope. We identify three profiles that reflect vendors' different positions toward AI vulnerabilities: proactive clarification (n = 46), silent (n = 115), and restrictive (n = 103). Our alignment results suggest that vendors may address AI vulnerability disclosure later than academic research and real-world incidents.

The Impact of Security and Privacy Controls on Users' Emotional Engagement with Generative AI Chatbots

Jabari Kwesi and Jiaxun Cao, Duke University; Hailee Cunningham, Duke Kunshan University; Pardis Emami-Naeini, Duke University

Available Media

Chatbots powered by generative AI (e.g., OpenAI's ChatGPT and Google's Gemini) are increasingly being appropriated for emotional support and companionship. These tools offer a suite of security and privacy (S&amp;P) controls, including model training opt-outs and memory toggles, yet how the presence of these controls influences users' attitudes toward emotionally sensitive disclosure remains understudied. We conducted a mixed-methods vignette study with 354 U.S. participants to examine how S&amp;P controls influence users' willingness to engage with generative AI chatbots for emotional support, their perceptions of how protected they are when using these systems, and their perceptions of how effective the chatbots are for providing support. Controls enabling deletion of disclosures had the largest positive impact: these offerings outperformed technically sophisticated controls such as local-only processing and model training opt-outs, where participants expressed difficulty understanding the underlying mechanisms. Yet trust remains fragile, and participants often doubted S&amp;P controls would function as promised. We conclude with actionable recommendations informed by our results to bridge users' comprehension gaps, build credible assurances, and properly calibrate barriers for users in distress.

Assessing LLM Response Quality in the Context of Technology-Facilitated Abuse

Vijay Prakash, New York University; Majed Almansoori, University of Wisconsin–Madison; Donghan Hu, New York University; Rahul Chatterjee, University of Wisconsin–Madison; Danny Yuxing Huang, New York University

Available Media

Technology-facilitated abuse (TFA) is a pervasive form of intimate partner violence (IPV) that leverages digital tools to control, surveil, or harm survivors. While tech clinics are one of the reliable sources of support for TFA survivors, they face limitations due to staffing constraints and logistical barriers. As a result, many survivors turn to online resources for assistance. With the growing accessibility and popularity of large language models (LLMs), and increasing interest from IPV organizations, survivors may begin to consult LLM-based chatbots before seeking help from tech clinics.

In this work, we present the first expert-led manual evaluation of four LLMs—two widely used general-purpose non-reasoning models and two domain-specific models designed for IPV contexts—focused on their effectiveness in responding to TFA-related questions. Using real-world questions collected from literature and online forums, we assess the quality of zero-shot single-turn LLM responses generated with a survivor safety-centered prompt on criteria tailored to the TFA domain. Additionally, we conducted a user study to evaluate the perceived actionability of these responses from the perspective of individuals who have experienced TFA.

Our findings, grounded in both expert assessment and user feedback, provide insights into the current capabilities and limitations of LLMs in the TFA context and may inform the design, development, and fine-tuning of future models for this domain. We conclude with concrete recommendations to improve LLM performance for survivor support.

A Large-Scale Study of Personalized Phishing using Large Language Models

Stefan Czybik, BIFOLD & TU Berlin; Anne Josiane Kouam, Inria & TU Berlin; Peter Heubl and Jan Magnus Nold, Ruhr-University Bochum; Konrad Rieck, BIFOLD & TU Berlin

Available Media

Large Language Models (LLMs) can generate fluent and persuasive text, making them valuable tools for communication. However, this capability also renders them attractive for malicious purposes. While several studies have shown that LLMs can support generic phishing, their potential for personalized attacks at scale has not been explored and quantified yet. In this study, we thus evaluate the effectiveness of LLM-based spear phishing in an experiment with 7700 participants. Using the target email addresses as queries, we collect personal information through web searches and automatically generate emails tailored to each participant. Our findings reveal a concerning situation: LLM-based spear phishing almost triples the click rate compared to generic phishing strategies. This effect is consistent, regardless of whether the generic emails are written by humans or generated by LLMs as well. Moreover, the cost of personalization is minimal, with approximately $0.03 per email. Given that phishing is still a major attack vector against IT infrastructures, we conclude that there is a pressing need to strengthen existing defenses, for example, by limiting publicly available information linkable to email addresses and incorporating personalized phishing into awareness trainings.

What Users Ask, Policies Miss: Unveiling the Gap Between Community-Expressed Privacy Concerns and LLM Provider Policies

Zhihuang Liu, Zhen Huang, Ling Hu, Yifan Yang, and Zhiping Cai, National University of Defense Technology

Available Media

Large Language Models(LLMs) process millions of conversations containing sensitive information daily, yet whether their privacy policies adequately address users' publicly expressed concerns remains unexplored. This paper presents the first large-scale, user-centered audit of privacy policy adequacy in LLM services. We systematically extract privacy concerns from Reddit communities of five major LLM providers and assess whether their latest privacy policies address these concerns. Our semi-automated pipeline analyzes 1,531 threads to extract 4,994 authentic privacy concerns, which we organize into a 20-topic taxonomy spanning four thematic groups. We then identify 3,137 policy gap instances and classify them into six categories: four policy coverage gaps (detail vague, AI feature unaddressed, vulnerable group neglected, and jurisdiction unclear) and two user perception gaps (explicit distrust and awareness deficit). Our analysis reveals that coverage gaps and perception gaps contribute nearly equally (50.2% vs. 49.8%), with AI-specific feature gaps (36.6%) and user awareness deficits (40.7%) together comprising the vast majority (77.3%) of all identified gaps. Notably, we also surface user-reported evidence suggesting potential discrepancies between stated policies and observed system behavior, highlighting the need for verifiable privacy guarantees. These findings demonstrate that improving LLM privacy requires dual-pronged interventions addressing both inadequate policy disclosures and user comprehension barriers, offering actionable insights for relevant stakeholders.

"Do Not Mention This to the User": Detecting and Understanding Malicious Agent Skills in the Wild

Yi Liu, Zhihao Chen, and Yanjun Zhang, Griffith University; Gelei Deng, Nanyang Technological University; Yuekang Li, University of New South Wales; Jianting Ning, Zhejiang Key Laboratory of Digital Fashion and Data Governance, Zhejiang Sci-Tech University; Leo Yu Zhang, Griffith University

Available Media

LLM-based coding agents increasingly rely on third-party extensions called skills, which bundle natural language instructions and helper scripts that execute with full user privileges. Community registries have emerged to distribute these skills, but the security implications remain unstudied due to the absence of labeled threat data. This paper presents a systematic security analysis of 98,380 skills collected from two major registries. Through a combination of static pattern matching and dynamic behavioral verification, we identify 157 skills exhibiting confirmed malicious behavior, encompassing 632 distinct vulnerabilities across 13 attack techniques. Our analysis reveals that these threats are deliberate rather than accidental: each malicious skill contains an average of 4.03 vulnerabilities spanning multiple attack phases. We identify two dominant attack strategies with statistically significant negative correlation---credential theft via remote code execution, and agent manipulation through adversarial instructions embedded in documentation. Over half of all confirmed cases originate from a single threat actor employing templated brand impersonation at scale. We further observe that attack sophistication correlates with concealment investment, with advanced skills universally employing undocumented capabilities while also exploiting platform-native trust mechanisms. Following responsible disclosure, registry maintainers removed all 157 (100%) of the reported skills. Our dataset and detection pipeline are publicly available to facilitate future research on securing LLM agent ecosystems.

Track 4

Software Security 1

Session Chair: Tianchang Yang, The Pennsylvania State University

Grand Ballroom I-IV

Exposing Resource-Exhaustion DoS Vulnerabilities with Leak-Oriented Minimum Path Covers

Lige Zhan and Yafei He, Wuhan University; Jiang Ming, Tulane University; Guojun Peng and Jianming Fu, Wuhan University

Available Media

Resource leaks---failures to properly release scarce Garbage Collection (GC)-unmanaged resources such as file descriptors, sockets, and database connections---are a pervasive root cause of resource-exhaustion denial-of-service (DoS) risks in Java programs. Prior approaches typically enumerate execution paths and inspect them to identify resources that are not properly released. However, exceptions can trigger non-local control transfers that bypass resource releases, and determining which such exceptions are leak-relevant requires context-sensitive reasoning beyond simple heuristics. Worse still, exhaustive path enumeration may trigger path explosion, significantly reducing the practicality of existing methods.

Therefore, we develop LeakHunter, a novel Java resource leak detector. LeakHunter first leverages an LLM to infer leak-relevant exceptions in context and statically validates them. Next, we propose a Leak-oriented Minimum Path Cover (LMPC) technique. Our insight is that there is no need to enumerate all paths: since (1) leak detection is existential, a single leak-prone path suffices to demonstrate that a resource-managing object is leak-prone, and (2) we can cover the resource-relevant code regions by leveraging a minimum path cover. Therefore, we aim to select a representative path set generated from MPC while being biased toward leak-prone behaviors. Accordingly, we construct witness paths from multiple MPCs and prioritize those with more acquisitions and fewer releases as the LMPC, as they are more likely to expose leaks. Across two public benchmarks, LeakHunter achieves the best detection performance while providing an 11.4x speedup in path analysis over the latest approach. In our real-world study, LeakHunter identifies 67 true leaks with potential DoS impact, including 15 confirmed by developers.

Liquidity Mining as an Attack Surface: Incentive-Induced Liquidity Attacks in Concentrated Liquidity Market Makers

Nora Sinong Lu and Chon Kit Lao, University of Macau; Yunlong Mao, Nanjing University; Xiaobo Zhou, University of Macau; Ruizhe Jia, Stanford University; Kanye Ye Wang, University of Macau

Available Media

Decentralized finance (DeFi) platforms commonly deploy volume-based incentive programs to bootstrap liquidity and improve market efficiency. We show that such subsidies, while well-intentioned, can create an underexplored economic attack surface in concentrated liquidity market makers (CLMMs). These subsidies enable a novel manipulation strategy. Attackers attract subsidized trading volume by providing highly concentrated liquidity, artificially inflating yields and luring victim liquidity providers into narrow price ranges. Once sufficient victim liquidity has accumulated, the attacker induces a price crash to extract value. We formalize this behavior as an incentive-induced liquidity attack and prove that it arises endogenously whenever subsidy intensity and volume caps exceed critical thresholds. Although subsidies are intended to bootstrap liquidity, these attacks ultimately lead to substantial liquidity depletion, causing platforms to incur incentive costs while market liquidity deteriorates. We conduct a large-scale measurement study on incentivized CLMM pools on BSC that are tied to Binance Alpha programs. We identify 29 attack events across 18 pools with 99 positive-loss victim addresses across 13 loss-bearing episodes and $2.87M in measured victim loss. These measurements show that incentive programs can amplify manipulation by introducing externalities that enable new extraction opportunities. Based on our findings, we propose mitigation strategies that reshape incentive parameters and information disclosure to reduce liquidity-provider exposure to incentive-induced attacks. Our findings show that incentive mechanisms designed to improve market quality can instead create the security vulnerabilities they aim to prevent.

Lost in Blockchain Address Misuse: Hidden Cross-Platform Risks and Their Security Impact

Zhenzhe Shao, Sun Yat-sen University and Zhejiang University; Jiashuo Zhang, Peking University; Zihao Li, University of Electronic Science and Technology of China and The Hong Kong Polytechnic University; Daoyuan Wu, Lingnan University; Chong Chen and Yiming Shen, Sun Yat-sen University; Lingfeng Bao, Zhejiang University and Hangzhou High-Tech Zone (Binjiang) Institute of Blockchain and Data Security; Yanlin Wang, Sun Yat-sen University; Jiachi Chen, Zhejiang University and Hangzhou High-Tech Zone (Binjiang) Institute of Blockchain and Data Security

Available Media

Blockchain systems, such as Ethereum, employ an account-based model, where each account is uniquely identified by an address. As the fundamental interface for user interaction and asset security, addresses are critical but also pose significant risks when misused. In this paper, we systematically reveal and analyze a class of risks termed Address Misuse, which includes two categories: Contract Account (CA) Misuse and Externally Owned Account (EOA) Misuse. Specifically, CA Misuse arises when users mistakenly treat non-contract addresses (NCAs) as CAs, while EOA Misuse occurs when users interact with EOAs whose private keys are exposed. For each category, we reveal the underlying mechanisms and also introduce previously undisclosed attack vectors that enable attackers to exploit these vulnerabilities for profit. To evaluate their prevalence and impacts, we first construct a dataset from GitHub and Stack Exchange, which contains addresses of various blockchain networks. This dataset includes 10 million candidate addresses for misuse analysis and 16 million exposed private keys. We then perform a large-scale on-chain analysis of their associated transactions on Ethereum and BSC. By combining heuristic rules, transaction pattern analysis, and symbolic execution, we identify 65,340 high-risk address instances, with associated asset losses amounting to about 127k ETH and 17.7k BNB, equivalent to over $574.8M. We evaluate the accuracy of our detection methods to ensure the reliability of the results, achieving an overall precision of 99.11%. Besides, our empirical evaluation also reveals two novel, previously undisclosed attack vectors, providing real-world evidence of how attackers actively exploit users' address misuse for profit.

Arguzz: Testing zkVMs for Soundness and Completeness Bugs

Christoph Hochrainer, Technische Universität Wien; Valentin Wüstholz, Diligence Security; Maria Christakis, Technische Universität Wien

Available Media

Zero-knowledge virtual machines (zkVMs) are increasingly deployed in decentralized applications and blockchain rollups since they enable verifiable off-chain computation. These VMs execute general-purpose programs, frequently written in Rust, and produce succinct cryptographic proofs. However, zkVMs are complex, and bugs in their constraint systems or execution logic can cause critical soundness (accepting invalid executions) or completeness (rejecting valid ones) issues.

We present ARGUZZ, the first automated tool for testing zkVMs for soundness and completeness bugs. To detect such bugs, ARGUZZ combines a novel variant of metamorphic testing with fault injection. In particular, it generates semantically equivalent program pairs, merges them into a single Rust program with a known output, and runs it inside a zkVM. By injecting faults into the VM, ARGUZZ mimics malicious or buggy provers to uncover overly weak constraints.

We used ARGUZZ to test six real-world zkVMs—RISC Zero, Nexus, Jolt, SP1, OpenVM, and Pico—and found eleven bugs in three of them. One RISC Zero bug resulted in a $50,000 bounty, despite prior audits, demonstrating the critical need for systematic testing of zkVMs.

Cutting the Gordian Knot: Detecting Malicious PyPI Packages via a Knowledge-Mining Framework

Wenbo Guo, Nanyang Technological University; Chengwei Liu, Nankai University; Ming Kang, Sichuan University; Yiran Zhang and Jiahui Wu, Nanyang Technological University; Zhengzi Xu, Imperial Global Singapore; Vinay Sachidananda and Yang Liu, Nanyang Technological University

Available Media

The Python Package Index (PyPI) has become a target for malicious actors, yet existing detection tools generate false positive rates of 15-30%, incorrectly flagging one-third of legitimate packages as malicious. This problem arises because current tools rely on simple syntactic rules rather than semantic understanding, failing to distinguish between identical API calls serving legitimate versus malicious purposes. To solve this challenge, we propose PyGuard, a knowledge-driven framework that converts detection failures into useful behavioral knowledge by extracting patterns from existing tools' false positives and negatives. Our method uses hierarchical pattern mining to identify behavioral sequences that separate malicious from benign code, employs Large Language Models to create semantic abstractions beyond syntactic variations, and combines this knowledge into a detection system that merges exact patterns matching with contextual reasoning. PyGuard achieves 99.50% accuracy with only 2 false positives versus 1,927-2,117 in existing tools, maintains 98.28% accuracy on obfuscated code, and identified 219 previously unknown malicious packages in real-world deployment. The behavioral patterns show cross-ecosystem applicability with 98.07% accuracy on NPM packages, demonstrating that semantic understanding enables knowledge transfer across programming languages.

Patch-Guided Vulnerability Detection: Extracting Java API Security Rules via Attack–Defense Cross-Analysis

Bofei Chen, Shuang Liao, and Lei Zhang, Fudan University; Chibin Zhang and Mathias Payer, EPFL; Yuan Zhang, Fudan University

Available Media

Security-sensitive APIs are critical components in modern Java applications, yet improper usage of these APIs frequently leads to severe vulnerabilities such as remote code execution. Existing methods for generating API security rules are limited as they rely on incomplete documentation or infer patterns from source code based on discovered inconsistencies.

We introduce VulGenie, a patch-driven framework that extracts precise API security rules from confirmed security patches to then detect API misuse vulnerabilities. VulGenie addresses three key challenges. First, it isolates violated constraints and defenses-related changes from noisy patches using our novel modification behavior dependency patch graph datastructure. Second, it identifies protected security-sensitive APIs and synthesizes rules through attack-defense cross-validation. Third, it scales analysis with adaptive, deviation-guided static analysis to balance precision and performance. Evaluated on 150 recent Java security patches, VulGenie extracts 198 API security rules with 81.82% precision, uncovering 177 rules absent in CodeQL. On ten popular Java applications, VulGenie detects 46 0-day vulnerabilities, substantially outperforming state-of-the-art works. Through our responsible vulnerability disclosure, 25 vulnerabilities have already been fixed with ten CVE identifiers assigned.

Regular Expression Denial of Service Induced by Backreferences

Yichen Liu, Stony Brook University; Berk Çakar, Purdue University; Aman Agrawal and Minseok Seo, Stony Brook University; James C. Davis, Purdue University; Dongyoon Lee, Stony Brook University

Available Media

This paper presents the a systematic and theoretical study of denial-of-service vulnerabilities in Regular Expressions with Backreferences (REwB). We introduce the Two-Phase Memory Automaton (2PMFA), an automaton model that precisely captures REwB semantics. Using this model, we derive necessary conditions under which backreferences induce super-linear backtracking runtime, even when sink ambiguity is linear, a regime where existing detectors report no vulnerability. Based on these conditions, we identify three vulnerability patterns and validate them in practice. Using the Snort intrusion detection ruleset, our evaluation identifies 48 previously unknown REwB vulnerabilities with quadratic or worse runtime. We further demonstrate practical exploits against Snort, including slowing rule evaluation by 0.6-1.2 seconds and bypassing alerts by triggering PCRE's matching limit.

Revealing the Dark Side of Smart Accounts: An Empirical Study of EIP-7702 Incurred Risks in Blockchain Ecosystem

Mingyuan Huang, Hong Kong University of Science and Technology; Han Liu, College of Cryptology and Cyber Science, Nankai University; Shuo Yang, Sun Yat-sen University; Daoyuan Wu, Lingnan University; Shuai Wang, Hong Kong University of Science and Technology

Available Media

The introduction of smart accounts by EIP-7702 represents a major advancement for blockchain account abstraction, enabling externally owned accounts (EOAs) to be upgraded into programmable accounts while still preserving their original addresses. This advancement significantly enhances both account functionality and usability, but also redefines blockchain trust boundaries between EOAs and smart contract accounts (CAs), thereby altering security assumptions and creating opportunities for novel types of attack.

To systematically examine these risks, we classify smart account-based risks into three categories according to the type of victim accounts: EOA-targeted, CA-targeted, and composite attacks. We then develop specialized detection tools that combine large-scale transaction analysis with cross-contract static analysis to identify malicious behaviors. Applying these tools across seven blockchains that support EIP-7702, we detect 924 malicious contract accounts, including several previously unreported zero-day cases. These attacks have led to more than 2.3 million in losses and exposed over 10 million to potential compromise. We uncover multiple key insights into attacker behaviors. Specifically, we find that over 63% of EIP-7702 authorization transactions are associated with malicious EOA-targeted attacks, and nearly half of the most frequently authorized contracts are controlled by attackers. In addition, we identify existing evasion tactics that attackers use to circumvent detection, attack impacts observed in real-world incidents, and potential risks that may emerge in future deployments, underscoring the urgency of addressing smart account security in blockchain ecosystems.

Track 5

Enigma Track: AI Vulnerabilities

Session Chair: Kevin Fu, Northeastern University

Dover Room

Improving Safety with Privacy: An Exploration of Privacy-Enhancing Safety Guardrails for Cloud Inference

Sai Deep Tetali, OpenAI

Available Media

Cloud AI systems face two obligations that are often framed as a tradeoff. Providers need to prevent misuse and investigate failures, while users and enterprises need assurances on how their data is stored, accessed, and processed. This framing assumes that safety requires broad visibility into user data and that privacy requires giving up meaningful safeguards.

This talk argues that privacy-enhancing infrastructure can improve both. We will explore how encryption, confidential computing, remote attestation, and other technologies can provide strong technical guarantees about how data is accessed and processed. These mechanisms can support strong safety guardrails without making routine human access the default. They provide stronger, auditable guarantees than from policy alone. We will also present a number of challenges with these architectures, including trust boundaries, operational tradeoffs, appeals, incidents, and software updates.

This is an open area of exploration and we are actively seeking to work with the community to set the standards for how responsible AI can be built.

Sai Deep Tetali is a Member of Technical Staff at OpenAI, where he leads the Private Stack team. His work focuses on building privacy-preserving infrastructure for frontier AI. Before OpenAI, Sai worked at Meta on privacy for augmented- and virtual-reality devices, spanning privacy-aware infrastructure and confidential computing. Earlier, he built machine-learning systems for detecting Android malware at Google and worked on automated software verification at Microsoft Research. He holds a PhD in Computer Science from UCLA, where his research focused on secure cloud computing, and is a co-author of The Android Malware Handbook.

Detecting Security Posture Drift in AI-Driven Software Development via Continuous Threat Modeling

Preeti Ravindra

Available Media

AI-assisted software development is accelerating how quickly systems are designed, modified, and deployed. While modern security tooling can identify vulnerabilities, misconfigurations, and policy violations, it often struggles to answer a more fundamental question: does the deployed system still reflect the intended security posture?

This talk introduces an approach for automatically deriving threat models from code, infrastructure, and configuration artifacts and tracking their evolution over time. Rather than focusing on code diffs, the approach computes threat model diffs to detect security intent drift, including changes to attack surfaces, trust boundaries, privilege relationships, and sensitive data flows. Through real-world examples of AI-assisted infrastructure changes, I show how security-relevant behavior can emerge without introducing traditional vulnerabilities, and how threat model differencing can surface these risks. The result is a practical workflow for maintaining security oversight in development environments where code changes faster than humans can review it.

Preeti is a technical leader who has delivered security outcomes by applying AI to security problems and ensuring AI systems are reliable and trustworthy for over a decade. Her career spans startups to Fortune 100 companies where she has advanced research and implemented it into scalable, revenue-aligned systems. She solves problems in both product security and security operations bridging research and engineering execution with equal ease. Her value proposition is working across functions, cutting through ambiguity and delivering measurable business value.

As a recognized industry voice, Preeti speaks at conferences like DEFCON,BSides and serves on program committees for CAMLIS, unprompted. She is passionate about bridging the AI and security communities and advancing early career professionals to more senior roles. Currently, Preeti is pursuing AI safety research in AI control.

5:00 pm–6:00 pm

Poster Session and Happy Hour

Harborside Ballroom

Check out the cool new ideas and the latest preliminary work on display at the Poster Session and Happy Hour. Take advantage of the opportunity to mingle with colleagues who may share your area of interest while enjoying complimentary food and drinks.

6:00 pm–7:00 pm

Open Forum with the Program Co-Chairs

Grand Ballroom V

Thursday, August 13

8:00 am–9:00 am

Continental Breakfast

Grand Foyer North and West

9:00 am–10:30 am

Track 1

Privacy Anonymity 2

Session Chair: Eyal Ronen, Tel Aviv University

Grand Ballroom VII-X

Breaking Widely Deployed Perceptual Hash Functions: Black-Box Collisions in Apple NeuralHash and Microsoft PhotoDNA

Diane Leblanc-Albarel and Bart Preneel, KU Leuven

Available Media

Perceptual hash functions have been designed to detect multimedia copyright violations and illegal content. To achieve their purpose, they map inputs that are perceived as similar to close outputs. For many widely deployed schemes, however, both the design strategy and detailed specifications remain proprietary. Governments are now considering their extension to Client-Side Scanning (CSS) for end-to-end encrypted services, verifying content against illegal material before encryption. In 2021, Apple presented a detailed proposal for CSS based on the NeuralHash perceptual hash function. After strong criticism over privacy and security concerns, Apple withdrew the proposal, but NeuralHash remains deployed on all devices, with its current purpose undisclosed.\ In theory, brute-force collisions for NeuralHash (96-bit hash value) require 248 evaluations. Shortly after the NeuralHash release, researchers showed it is easy to craft perceptually dissimilar collisions, to incriminate any user by sending an innocent image sharing the same hash value as illegal content. This work shows a more serious weakness: when inputs are restricted to human faces, we found several collisions between perceptually different images after only 216 hash function evaluations. Unlike targeted attacks, our black-box approach requires no knowledge of the hash function design. We also demonstrate a high false negative rate (images that should share the same hash but do not). We further confirm the generality of our approach by studying PhotoDNA, Microsoft's widely deployed 1152-bit perceptual hash function. In the case of PhotoDNA, we found near-collisions at thresholds significantly lower than previously reported, appearing after between 214.6 and 217 evaluations depending on the threshold used. This is the first work to demonstrate exact collisions in NeuralHash and to identify near-collisions in PhotoDNA at such low thresholds. These results cast serious doubts on the suitability of these designs for large-scale client scanning, as they produce high false positive and false negative rates, and highlight the need to reassess their security and feasibility, particularly for large-scale applications where privacy risks and false positives have serious consequences.

SMASH: Scalable Maliciously Secure Hybrid Multi-party Computation Framework for Privacy-Preserving Large Language Models

Yunlv Lv and Rui Zhang, Institute of Information Engineering, Chinese Academy of Sciences; State Key Laboratory of Cyberspace Security Defense; School of Cyber Security, University of Chinese Academy of Sciences; Zhiyuan Zhang, Max Planck Institute for Security and Privacy; Ziyi Wan, Institute of Information Engineering, Chinese Academy of Sciences; State Key Laboratory of Cyberspace Security Defense; School of Cyber Security, University of Chinese Academy of Sciences; Lanxue Zhang, Institute of Information Engineering, Chinese Academy of Sciences; School of Cyber Security, University of Chinese Academy of Sciences; Minhui Xue, CSIRO's Data61 and Responsible AI Research (RAIR) Centre, Adelaide University; Jiangtao Li, East China Normal University; Yanan Cao, Institute of Information Engineering, Chinese Academy of Sciences; School of Cyber Security, University of Chinese Academy of Sciences

Available Media

The meteoric rise of Large Language Models (LLMs) has sparked an urgent need for privacy-preserving inference. However, existing maliciously secure multi-party computation (MPC) frameworks face a "performance collapse" when scaling to large models, primarily due to the quadratic (O(n2)) communication overhead of nonlinear operators and expensive share conversions. This paper presents SMASH, a highly scalable, maliciously secure hybrid MPC framework that shatters these bottlenecks. SMASH introduces a novel DFT-based rotation technique and a lightweight zero-knowledge proof of knowledge (ZKPoK) construction to evaluate nonlinear operations. For the first time, this approach achieves linear communication complexity (O(n)) relative to the party count, independent of function complexity. Furthermore, SMASH provides a suite of high-efficiency conversion protocols (A2L/L2A and SM-LUT-based A2B/B2A) that bridge arithmetic and Boolean domains without relying on costly cryptographic primitives. Extensive benchmarks demonstrate that SMASH outperforms state-of-the-art frameworks (e.g., MP-SPDZ, MD-ML) by up to 18.9× in runtime and achieves a communication reduction of up to 103×. With its constant-round online phase and low WAN sensitivity, SMASH paves the way for secure, geographically distributed LLM deployments, achieving an unprecedented balance between adversarial robustness and practical efficiency.

Large-scale online deanonymization with LLMs

Simon Lermen, MATS; Daniel Paleka, Joshua Swanson, and Michael Aerni, ETH Zurich; Nicholas Carlini, Anthropic; Florian Tramèr, ETH Zurich

Available Media

We show that large language models can be used to perform at-scale deanonymization. With full Internet access, our agent can re-identify Hacker News users and Anthropic Interviewer participants at high precision, given pseudonymous online profiles and conversations alone, matching what would take hours for a dedicated human investigator. We then design attacks for the closed-world setting. Given two databases of pseudonymous individuals, each containing unstructured text written by or about that individual, we implement a scalable attack pipeline that uses LLMs to: (1) extract identity-relevant features, (2) search for candidate matches via semantic embeddings, and (3) reason over top candidates to verify matches and reduce false positives. Compared to classical deanonymization work (e.g., on the Netflix prize) that required structured data, our approach works directly on raw user content across arbitrary platforms. We construct three datasets with known ground-truth data to evaluate our attacks. The first links Hacker News to LinkedIn profiles, using cross-platform references that appear in the profiles. Our second dataset matches users across Reddit movie discussion communities; and the third splits a single user's Reddit history in time to create two pseudonymous profiles to be matched. In each setting, LLM-based methods substantially outperform classical baselines, achieving up to 55% recall at 90% precision compared to near 0% for the best non-LLM method. Our results show that the cost of deanonymizing pseudonymous users online has fallen sharply, and threat models for online privacy should account for this shift.

United We Defend: Collaborative Membership Inference Defenses in Federated Learning

Li Bai, The Hong Kong Polytechnic University; Junxu Liu, The Hong Kong Polytechnic University and PolyU Research Centre for Privacy and Security Technologies in Future Smart Systems; Sen Zhang, Xinwei Zhang, and Qingqing Ye, The Hong Kong Polytechnic University; Haibo Hu, The Hong Kong Polytechnic University and PolyU Research Centre for Privacy and Security Technologies in Future Smart Systems

Available Media

Membership inference attacks (MIAs), which determine whether a specific data point was included in the training set of a target model, have posed severe threats in federated learning (FL). Unfortunately, existing MIA defenses, typically applied independently to each client in FL, are ineffective against powerful trajectory-based MIAs that exploit temporal information throughout the training process to infer membership status. In this paper, we investigate a new FL defense scenario driven by heterogeneous privacy needs and privacy-utility trade-offs, where only a subset of clients are defended, as well as a collaborative defense mode where clients cooperate to mitigate membership privacy leakage. To this end, we introduce CoFedMID, a collaborative defense framework against MIAs in FL, which limits local model memorization of training samples and, through a defender coalition, enhances privacy protection and model utility. Specifically, CoFedMID consists of three modules: a class-guided partition module for selective local training samples, a utility-aware compensation module to recycle contributive samples and prevent their overconfidence, and an aggregation-neutral perturbation module that injects noise for cancellation at the coalition level into client updates. Extensive experiments on three datasets show that our defense framework significantly reduces the performance of seven MIAs while incurring only a small utility loss. These results are consistently verified across various defense settings.

Analyzing Cryptography in Context: A Cryptography-Native Approach to Threat Modeling

Ran Canetti and Julie Ha, Boston University; Gabriel Kaptchuk, University of Maryland, College Park

Available Media

We observe that the existing norms within cryptography do not expect protocol analysts to document the sociotechnical properties that a deployed system should have. To help close this potential gap, we develop a framework that allows bringing sociotechnical dimensions into analyses of cryptographic systems, and in particular facilitates "in-context" analysis of proposed cryptographic deployments on top of widely-accepted cryptographic modeling techniques. To explore the utility of our framework, we use Apple's 2021 CSAM scanning proposal as a case study. We show how our framework naturally surfaces many of the criticisms of Apple's proposal and helps us identify a previously undocumented property of the proposal.

PVMark: Enabling Public Verifiability for LLM Watermarking Schemes

Haohua Duan, East China University of Science and Technology and Shanghai Jiao Tong University; Liyao Xiang, Shanghai Jiao Tong University and Shanghai Innovation Institute; Xin Zhang, Ant Group; Baochun Li, University of Toronto; Bo Li, Hong Kong University of Science and Technology

Available Media

Watermarking schemes for large language models (LLMs) have been proposed to identify the source of the generated text. However, current watermarking solutions hardly resolve the trust issue: the watermark detection often relies on the secret key and thus remains opaque to the public; otherwise any adversary may launch removal attacks if the secret key is exposed. To resolve the dilemma, we propose PVMark, a plugin based on zero-knowledge proof (ZKP), enabling the watermark detection process to be publicly verifiable by third parties without disclosing any secret key. PVMark novelly hinges upon the proof of `correct execution' of watermark detection on which a set of constraints are built, and is optimized according to the structural nature of the detection process. Developed for three representative watermarking schemes, we implement multiple variants of PVMark in Python, Rust and Circom, covering combinations of three hash functions and four ZKP protocols, showing our approach effectively works under a variety of circumstances. By experimental results, PVMark efficiently enables public verifiability on the state-of-the-art LLM watermarking schemes yet without compromising the watermarking performance, and hence is promising for practical deployment.

Turn Your Face Into An Attack Surface: Screen Attack Using Facial Reflections in Video Conferencing

Yong Huang, Yanzhao Lu, Mingyang Chen, En Zhang, and Jiazi Li, Zhengzhou University; Wanqing Tu, Durham University

Available Media

In video conferencing, human faces serve as the primary visual focal points, playing multifaceted roles that enhance visual communication and emotional connection. However, we argue that a human face is also a side channel, which can unwittingly leak on-screen information through online video feeds. To demonstrate this, we conduct feasibility studies, which reveal that, illuminated by both ambient light and light emitted from displays, the human face can reflect optical variations of different on-screen content. The paper then proposes FaceTell, a novel side-channel attack system that eavesdrops on fine-grained application activities from pervasive yet subtle facial reflections during video conferencing. We implement FaceTell in a real-world testbed with three different brands of laptops and four mainstream video conferencing platforms. FaceTell is then evaluated with 24 human subjects across 13 unique indoor environments. With more than 12 hours of video data, FaceTell achieves a high accuracy of 99.32% for eavesdropping on 28 popular applications and is resilient to many practical impact factors. Finally, potential countermeasures are proposed to mitigate this new attack.

NOIR: Privacy-Preserving Generation of Code with Open-Source LLMs

Khoa Nguyen, Khiem Ton, and NhatHai Phan, New Jersey Institute of Technology; Issa Khalil, Hamad Bin Khalifa University; Khang Tran and Cristian Borcea, New Jersey Institute of Technology; Ruoming Jin, Kent State University; Abdallah Khreishah, New Jersey Institute of Technology; My T. Thai, University of Florida

Available Media

Although boosting software development performance, large language model (LLM)-powered code generation introduces intellectual property and data security risks rooted in the fact that a service provider (cloud) observes a client's prompts and generated code, which can be proprietary in commercial systems. To mitigate this problem, we propose NOIR, the first framework to protect the client's prompts and generated code from the cloud. NOIR uses an encoder and a decoder at the client to encode and send the prompts' embeddings to the cloud to get enriched embeddings from the LLM, which are then decoded to generate the code locally at the client. Since the cloud can use the embeddings to infer the prompt and the generated code, NOIR introduces a new mechanism to achieve indistinguishability, a local differential privacy protection at the token embedding level, in the vocabulary used in the prompts and code, and a data-independent and randomized tokenizer on the client side. These components effectively defend against reconstruction and frequency analysis attacks by an honest-but-curious cloud. Extensive analysis and results using open-source LLMs show that NOIR significantly outperforms existing baselines on benchmarks, including the Evalplus (MBPP and HumanEval, Pass@1 of 76.7 and 77.4), and BigCodeBench (Pass@1 of 38.7, only a 1.77% drop from the original LLM) under strong privacy against attacks.

Track 2

Security of ML 3

Session Chair: Omer Akgul, RSAC Labs

Grand Ballroom VI

Bypassing Prompt Guards in Production with Controlled-Release Prompting

Jaiden Fairoze, University of California, Berkeley; Sanjam Garg, University of California, Berkeley, and zkBricks Inc.; Keewoo Lee, Ethereum Foundation; Mingyuan Wang, New York University Shanghai

Available Media

Ball et al. recently established that prompt filtering for AI alignment faces a fundamental barrier: under standard cryptographic assumptions, no filter running significantly faster than the protected model can universally distinguish adversarial prompts from benign ones. We investigate whether this impossibility result translates to real-world vulnerabilities in deployed large language model (LLM) systems. We answer affirmatively by introducing controlled-release prompting, a practical instantiation of the theoretical framework that exploits the resource asymmetry between lightweight input filters and the main models they protect. Unlike the theoretical construction, our attack does not require model modification: it generates malicious prompts that are indecipherable by any bounded filter yet remain tractable to the target LLM. We find our attack to be successful on four major chat platforms (Google Gemini, DeepSeek Chat, xAI Grok, and Mistral Le Chat) where baseline methods fail. Additionally, we apply our attack to extract copyrighted data from Gemini. Finally, we provide a systematic evaluation of 14 open-weight prompt guard models, revealing that even reasoning-capable filters cannot reliably detect our attack without incurring prohibitive resource overhead.

What the Eyes See, the LLMs Miss: Exploiting Human Perception for Adversarial Text Attacks

Qin Yang, University of Connecticut; Lu Malloy, University of Tennessee; Joshua Lee, University of California, Santa Barbara; Xiaohan Chang, University of Connecticut; Meisam Mohammady, Iowa State University; Doowon Kim, University of Tennessee; Yuan Hong, University of Connecticut

Available Media

Large language model (LLM)-powered content moderation systems are a critical defense against harmful online content. However, they operate primarily on tokenized text and often overlook visual cues that humans naturally use when interpreting content. We show that this limitation creates a fundamental vulnerability: content readily recognized as harmful by humans can evade automated moderation. To systematically study this problem, we introduce Human-Perceptible Adversarial Attacks (HPAA), which embed harmful expressions into otherwise benign text using visually salient typographic manipulations. HPAA strategically combines features such as spacing, emphasis, and spatial arrangement to preserve human recognition while reducing machine detectability. Operating in a black-box setting with a small query budget, the attack automatically generates evasive content without model access or gradient information. We evaluate HPAA on multiple datasets and thirteen widely deployed moderation systems, including commercial APIs and state-of-the-art open-source guardrails. With only three detector queries, generated attacks achieve over 86% human recognition while keeping detection rates below 1% across evaluated systems. We further identify the typographic factors driving successful evasion, analyze why current moderation architectures fail to capture these signals, and discuss practical defenses. Our findings reveal a fundamental blind spot in current LLM-based moderation systems and motivate moderation approaches that better align with human perceptual understanding. Disclaimer: This paper includes examples of harmful, hateful, or abusive language for research purposes. Reader discretion is advised.

JailbreakScope: Interpreting Jailbreak Mechanism through Representation and Circuit Analyses

Zeqing He, Zhibo Wang, Zhixuan Chu, Huiyu Xu, Wenhui Zhang, Qinglong Wang, and Rui Zheng, Zhejiang University

Available Media

Large Language Models (LLMs) exhibit impressive performance but remain vulnerable to jailbreak attacks, where adversarial prompts are crafted to bypass safety alignments and elicit unexpected responses. Despite their prevalence, the underlying mechanisms that enable jailbreaks are still not well understood. Recent studies primarily focus on static representation shifts or on identifying components associated with generation safety. However, these studies neither explore diverse jailbreak patterns nor provide a fine-grained explanation from the failure of circuit to representation changes, leaving significant gaps in uncovering jailbreak mechanism. In this paper, we propose JailbreakScope, an interpretation framework that analyzes jailbreak mechanisms from both representation (how jailbreaks distort LLM's harmfulness perception) and circuit (how jailbreaks impact circuits that are important for generation safety) perspectives, tracking their evolution throughout the entire generation process. We conduct in-depth evaluations on 5 mainstream LLMs under 7 jailbreak strategies. Our evaluation reveals a general pattern that jailbreaks amplify components that reinforce affirmative responses while suppressing those producing refusal, which shifts representations towards safe regions, leading LLMs to provide responses instead of refusals. Moreover, we find a strong and consistent correlation between representation deception and circuit activation shift across diverse jailbreaks and multiple LLMs.

One Bad Token Spoils the Barrel: Assessment, Detection, and Remediation of Glitch Tokens in Large Language Models

Kunsheng Tang and Peigui Qi, University of Science and Technology of China; Yide Song, University of Washington; Wenbo Zhou, University of Science and Technology of China; Zhicong Huang, Ant Group; Qing Guo, Nankai University; Tianwei Zhang, Nanyang Technological University; Weiming Zhang and Nenghai Yu, University of Science and Technology of China; Jie Zhang, CFAR, A*STAR, Singapore

Available Media

Large Language Models (LLMs) have shown remarkable capabilities across numerous applications. However, the recent emergence of "glitch tokens," referring to tokens that cause unpredictable and erroneous model behaviors, poses significant reliability and security challenges. Despite prior investigations of glitch tokens, critical gaps remain, including insufficient safety assessments, limited detection coverage, and the absence of effective remediation strategies. To address these challenges, we first demonstrate that glitch tokens can readily bypass safety mechanisms and elicit unsafe outputs across LLMs through straightforward exploitation methods. More critically, we reveal that these tokens exhibit cross-model transferability, inducing safety risks across model families, tokenizers, commercial and moderation systems, underscoring their pervasive security implications. We then propose GlitchQuiz, a red-teaming framework inspired by human language acquisition, to comprehensively detect glitch tokens, surpassing existing detection methods. Finally, we develop GlitchEdit, a training-free embedding-layer editing approach that effectively remediates glitch tokens, reducing unsafe behaviors with average unsafe rates decreasing from 96.36% to 2.87% across evaluated LLMs and maintaining overall performance. Our findings have been responsibly disclosed to nine affected leading LLM providers, including OpenAI, Anthropic, and others, to help foster safer AI ecosystems.

Defending Jailbreak Attacks on Large Language Models via Manifold Trajectory Kinetics

Hangtao Zhang and Yucheng Zhao, Huazhong University of Science and Technology; Sishun Liu, Changsha University of Science and Technology; Ziqi Zhou, Huazhong University of Science and Technology; Zeyu Ye, Xiangtan University; Wei Wan, City University of Macau; Minghui Li and Shengshan Hu, Huazhong University of Science and Technology; Yanjun Zhang, Yi Liu, and Leo Yu Zhang, Griffith University

Available Media

Jailbreak prompts can bypass alignment guardrails in large language models (LLMs) and elicit unsafe outputs, making reliable deployment-time detection critical. Prior detection approaches largely rely on static detection of internal states (e.g., raw inputs, gradients, or hidden features). We show that such static signals may overlook how risk emerges and evolves during model processing, and can be brittle under two challenging cases: (i) pseudo-malicious prompts that are benign by intent but contain safety-related keywords, and (ii) adaptive attacks that explicitly optimize against the deployed detector. To overcome this limitation, we shift our perspective from static analysis of detection signals to analyzing their dynamic evolution along model depth, examining how the neighborhood structure of the underlying data manifold changes across layers. We present Manifold Trajectory Kinetics (MTK), which treats an LLM as a kinetic system transforming inputs into outputs and detects jailbreaks by tracking how a prompt's neighborhood structure evolves across layers. Benign prompts remain close to benign neighborhoods throughout inference, whereas jailbreak prompts exhibit a characteristic trajectory that begins near malicious seeds and later strategically shifts toward benign neighborhoods to evade refusal. Across four LLMs and ten jailbreak attacks, MTK achieves strong robustness to both failure modes: on pseudo-malicious prompts, it attains a jailbreak true positive rate of 95% at a false positive rate of 5% on benign prompts and 2% on pseudomalicious prompts, and under adaptive attacks, it maintains an average true positive rate of 85%. We further demonstrate the superior performance of MTK for jailbreak detection in vision-language models. Our code is available at https://github.com/Rookie143/mtk.

GateBreaker: Gate-Guided Attacks on Mixture-of-Expert LLMs

Lichao Wu, Sasha Behrouzi, and Mohamadreza Rostami, Technical University of Darmstadt; Stjepan Picek, University of Zagreb and Radboud University; Ahmad-Reza Sadeghi, Technical University of Darmstadt

Available Media

Mixture-of-Experts (MoE) architectures have advanced the scaling of Large Language Models (LLMs) by activating only a sparse subset of parameters per input, enabling state-of-the-art performance with reduced computational cost. As these models are increasingly deployed in critical domains, understanding and strengthening their alignment mechanisms is essential to prevent harmful outputs. However, existing LLM safety research has focused almost exclusively on dense architectures, leaving the unique safety properties of MoEs largely unexamined. The modular, sparsely-activated design of MoEs suggests that safety mechanisms may operate differently than in dense models, raising questions about their robustness.

In this paper, we present GateBreaker, the first training-free, lightweight, and architecture-agnostic attack framework that compromises the safety alignment of modern MoE LLMs at inference time. GateBreaker operates in three stages: (i) gate-level profiling, which identifies safety experts disproportionately routed on harmful inputs, (ii) expert-level localization, which localizes the safety structure within safety experts, and (iii) targeted safety removal, which disables the identified safety structure to compromise the safety alignment. Our study shows that MoE safety concentrates within a small subset of neurons coordinated by sparse routing. Selective disabling of these neurons, maximum 2.9% of neurons in the targeted expert layers, significantly increases the averaged attack success rate (ASR) from 7.4% to 64.9% against the eight latest aligned MoE LLMs with limited utility degradation. These safety neurons transfer across models within the same family, raising ASR from 17.9% to 67.7% with one-shot transfer attack. Furthermore, GateBreaker generalizes to five MoE vision language models (VLMs) with 60.9% ASR on unsafe image inputs. To our knowledge, no prior work achieves this level of efficacy against MoE LLMs.

From Length to Content: Token-Length Side-Channel Attacks on LLM API Merged Outputs

Sijia Li, Tianyu Cui, Miao Chen, Xinjie Lin, and Zheyuan Gu, Zhongguancun Laboratory; Xinhao Deng, Tsinghua University; Ke Xu and Qi Li, Tsinghua University and Zhongguancun Laboratory

Available Media

Large Language Models (LLMs) are increasingly accessed via remote APIs over encrypted channels. While previous studies have shown that fine-grained side channels can leak information under token-by-token processing, modern LLM services typically stream outputs in multi-token chunks, seemingly mitigating such token-level leakage. However, in this paper, we demonstrate that token aggregation does not eliminate privacy risks. Instead, it introduces a merged token-length side channel. In this channel, the sizes of encrypted response chunks inadvertently expose character-length patterns of merged token groups. To study the practical implications of this leakage, we propose PromptEcho, a passive eavesdropping attack that reconstructs semantically meaningful portions of natural-language responses from observations of merged token lengths. PromptEcho frames the reconstruction task as a constrained sequence recovery problem, leveraging semantics-aware reasoning to resolve the ambiguity caused by merged token transmission. Specifically, PromptEcho first applies semantics-aware reasoning to decompose merged-token length observations, then uses probabilistic semantic alignment with a fine-tuned language model's linguistic priors to infer the most likely underlying tokens.

We evaluate PromptEcho on real-world API interaction traffic from multiple commercial LLM services, correctly inferring the topic for 42.5% of sessions and reconstructing 25.5% of responses with a cosine similarity above 0.9 on OpenAI's GPT-4o and DeepSeek-V3. These findings demonstrate that merged-token transmission remains vulnerable to token-length side-channel attacks, indicating practical privacy risks from encrypted traffic observations under the evaluated API-based LLM settings.

When Memory Becomes a Vulnerability: Towards Multi-turn Jailbreak Attacks against Text-to-Image Generation Systems

Shiqian Zhao, Nanyang Technological University; Jiayang Liu, Nanyang Technological University; and Institute of Science Tokyo, Japan; Yiming Li, Runyi Hu, and Xiaojun Jia, Nanyang Technological University; Wenshu Fan, University of Electronic Science and Technology of China; Xiaobao Wu and Xinfeng Li, Nanyang Technological University; Jie Zhang, CFAR and IHPC, Agency for Science, Technology and Research (A*STAR), Singapore; Wei Dong and Tianwei Zhang, Nanyang Technological University; Luu Anh Tuan, Nanyang Technological University and VinUniversity

Available Media

Modern text-to-image (T2I) generation systems (e.g., DALL·E 3) exploit the memory mechanism, which captures key information in multi-turn interactions for faithful generation. Despite its practicality, the security analyses of this mechanism have fallen far behind. In this paper, we reveal that it can exacerbate the risk of jailbreak attacks. Previous attacks fuse the unsafe target prompt into one ultimate adversarial prompt, which can be easily detected or lead to the generation of non-unsafe images due to under- or over-detoxification. In contrast, we propose embedding the malice at the inception of the chat session in memory, addressing the above limitations.

Specifically, we propose Inception, the first multi-turn jailbreak attack against real-world text-to-image generation systems that explicitly exploits their memory mechanisms. Inception is composed of two key modules: segmentation and recursion. We introduce Segmentation, a semantic-preserving method that generates multi-round prompts. By leveraging NLP analysis techniques, we design policies to decompose a prompt, together with its malicious intent, according to sentence structure, thereby evading safety filters. Recursion further addresses the challenge posed by unsafe sub-prompts that cannot be separated through simple segmentation. It firstly expands the sub-prompt, then invokes segmentation recursively. To facilitate multi-turn adversarial prompts crafting, we build VisionFlow, an emulation T2I system that integrates two-stage safety filters and industrial-grade memory mechanisms. The experiment results show that Inception successfully allures unsafe image generation, surpassing the SOTA by a 20.0% margin in attack success rate. We also conduct experiments on the real-world commercial T2I generation platforms, further validating the threats of Inception in practice.

Track 3

Network Security 2

Session Chair: Xinwen Fu, University of Massachusetts Lowell

Grand Ballroom V

A Formal Security Analysis of CAN XL

Zhaozhou Tang, Georgia Institute of Technology; Khaled Serag, Qatar Computing Research Institute; Z. Berkay Celik, Purdue University; Vijay Ganesh, Saman Zonouz, and Raheem Beyah, Georgia Institute of Technology

Available Media

For decades, the Controller Area Network (CAN) has been the backbone of in-vehicle communication. As modern vehicles integrate cameras, LiDARs, and AI components, classic CAN (CAN CC) faces growing limitations in bandwidth, functionality, and security. To fill these gaps, CAN XL was introduced as the next generation of CAN, aiming to offer longer payloads, higher bandwidth, and enhanced security.

CAN XL makes significant standard-level changes across multiple stack layers. It also introduces several security features, but it is unclear whether these are mere add-on extensions or whether the standard redesign itself tackles CAN's chronic security weakness: the MAC sub-layer. This sub-layer governs frame formats and error handling and has historically enabled many CAN CC attacks. As the industry transitions to CAN XL, the security posture of its yet-unexplored MAC sub-layer must be understood before widespread deployment.

This paper presents the first security analysis of the CAN XL standard, focusing on its MAC sub-layer. We develop a bit-precise CAN XL formal model and release it to facilitate future research. We design a formal analysis workflow guided by CAN XL's field-oriented structure to uncover vulnerabilities. Contrary to expectations, our analysis shows that CAN XL remains vulnerable to all known CAN CC MAC sub-layer issues while introducing seven new vulnerabilities, arguably worsening security. We validate them using commercial CAN XL controllers and demonstrate exploitability via two multi-stage attacks on a testbed simulating real vehicle traffic. Finally, we propose mitigations including formally verifying standard revisions that could prevent several attacks.

Shielding QR Codes: Unveiling the Real-World Illicit Promotion Behind Adversarial QR Codes

Lijie Wu and Xiaoping Zhang, Tsinghua University; Mingxuan Liu, Zhongguancun Laboratory; Yue Qin, Central University of Finance and Economics; Baojun Liu, Tsinghua University; Geng Hong, Fudan University; Zhenrui Zhang and Chenghui Wu, Baidu Inc; Hui Jiang, Tsinghua University and Baidu Inc

Available Media

Adversarial QR Code Images (AQRIs) represent an emerging threat vector for covert (usually illicit) online promotion. They use adversarial perturbations to evade QR detectors (e.g., OCR-based models) while retaining decodability for information delivery, facilitating malicious content dissemination, and posing risks to both platforms and users. Though adversarial attacks are well-studied, targeted techniques against structured QR codes are underexplored.

To systematically investigate the real-world AQRI abuse, we cooperated with a leading Internet service provider. With the help of our partner, grounded in empirical observations, we introduce Adato, an enhanced framework for AQRI detection, by prioritizing finder pattern regions and identifying adversarial techniques through cross-platform consistency checks. Experimental results demonstrate that Adato achieves 98.6% precision and 97.8% recall on AQRI detection, significantly outperforming existing detectors. With the collaboration of our partner,we legally obtained posts with images from five well-known international social media platforms from September, 2024 to March, 2025, e.g., Reddit, Baidu Tieba. We applied Adato to over 40 million images and identified 68,467 AQRIs, demonstrating their widespread real-world use and their ability to evade existing moderation mechanisms. Analysis of our detected AQRIs reveals that AQRIs are widely used for illicit promotion: 95.78% are linked to 2,079 malicious URLs, spanning 7 business categories. Additionally, we analyzed the information dissemination strategies employed, such as redirect chains and indirect propagation paths that exploit cross-platform inconsistencies. These results highlight Adato's effectiveness in strengthening existing moderation and recognition pipelines against AQRI abuses.

Cloud-Native Carjacking: Fleet-wide Compromise via Telematics Authorization Failures

Yangyang Liu, The Hong Kong Polytechnic University; Zhengjie Du, Geely Automobile Research Institute; Xiaofang Wang, Geely University of China; Yang Yin, Geely Automobile Research Institute; Xiapu Luo, The Hong Kong Polytechnic University

Available Media

Modern connected vehicles increasingly rely on cloud-centric telematics backends to provide remote control, telemetry, media access, and other software-defined functionality. Although these platforms commonly employ endpoint hardening and encrypted communication, such measures do not ensure correct backend authorization. Across the platforms we evaluated, cloud services often accepted a valid vehicle identity without consistently binding that identity to the specific topic, object, or command target being accessed.

We investigate this failure mode by recovering reusable authentication material from a seed vehicle and testing whether backend authorization remains properly scoped to vehicle-specific resources across MQTT, HTTPS, and SMS. On several legacy platforms, process-memory inspection exposed client certificates, private keys, tokens, or session artifacts sufficient to authenticate to backend services. Using this material, we established authenticated backend sessions and then modified logical vehicle identifiers, such as Vehicle Identification Numbers (VINs), at the application layer. We discovered previously unknown authorization vulnerabilities in telematics platforms used by three major OEMs. Starting from a single seed vehicle under the attacker's control, we demonstrate cross-vehicle privilege escalation without physical access to victim vehicles and without OEM-internal privileges. These flaws enable remote vehicle control, unauthorized access to private surveillance media, and spoofed fallback commands. We conclude with root-cause analysis informed by coordinated disclosure and derive mitigation lessons for identity-bound, Zero-Trust-style telematics authorization.

DUPIN: Attack Learning Is Still Needed! Demonstrating Few-Shot after Unsupervised Pretraining Is A Nimble Forensics Learner

Chanwoo Bae, Purdue University; Hailun Ding, IBM Research; Shiqing Ma, University of Massachusetts Amherst; Xiangyu Zhang, Purdue University

Available Media

Advanced persistent threats (APTs) pose a significant challenge in cybersecurity, involving staged and prolonged operations that often remain undetected until postmortem indicators, such as sabotage or financial loss, emerge. In consequence, human analysts are facing a needle-in-a-haystack challenge among vast accumulation of daily audit logs. However, leveraging a learning system for attack forensics is limited due to the scarcity of attack data, as attacks occur infrequently. In addition, since malicious behaviors are usually embedded in massive benign ones, they are very hard to label by humans. Thus, the recent approaches leverage self-supervised learning methods, where models rely solely on benign data and perform outlier detection. However, these methods struggle with the increasing complexity and dynamics of large-scale audit logs, often resulting in non-trivial false positives. Therefore, we propose a novel approach to learning-based attack forensics called DUPIN. First, DUPIN performs unsupervised pre-training on an enormous amount of audit events in the form of provenance graphs. It then proceeds to a few-shot learning stage, leveraging a small number of labeled attack examples to fine-tune its detection capabilities. We pretrain DUPIN on up to 38 - 52 days of audit logs (7.3TB total) and evaluate it against various baselines on 25 APT campaigns across four different data sources, facilitating the scalable evaluation.

SoK: The Pitfalls of Deep Reinforcement Learning for Cybersecurity

Shae McFadden, King's College London, The Alan Turing Institute, and University College London; Myles Foley, The Alan Turing Institute and Devotion AI Labs; Elizabeth Bates, The Alan Turing Institute; Ilias Tsingenopoulos, KU Leuven; Sanyam Vyas, Cardiff University and The Alan Turing Institute; Vasilios Mavroudis, King's College London; Chris Hicks, The Alan Turing Institute; Fabio Pierazzi, University College London

Available Media

Deep Reinforcement Learning (DRL) has achieved remarkable success in domains requiring sequential decision-making, motivating its application to cybersecurity problems. However, transitioning DRL from laboratory simulations to bespoke cyber environments can introduce numerous issues. This is further exacerbated by the often adversarial, non-stationary, and partially-observable nature of most cybersecurity tasks. In this paper, we identify and systematize 11 methodological pitfalls that frequently occur in DRL for cybersecurity (DRL4Sec) literature across the stages of environment modeling, agent training, performance evaluation, and system deployment. By analyzing 66 significant DRL4Sec papers (2018-2025), we quantify the prevalence of each pitfall and find an average of over five pitfalls per paper. We demonstrate the practical impact of these pitfalls using controlled experiments in (i) autonomous cyber defense, (ii) adversarial malware creation, and (iii) web security testing environments. Finally, we provide actionable recommendations for each pitfall to support the development of more rigorous and deployable DRL-based security systems.

Cutting the Fuse: Actionable APT Attack Blocking in Provenance-based IDS

Weiheng Wu, School of Cyber Engineering, Xidian University; Wei Qiao, School of Cyber Engineering, Xidian University; State Key Laboratory of Integrated Services Networks (ISN); Teng Li, School of Cyber Engineering, Xidian University; State Key Laboratory of Integrated Services Networks (ISN); Songshan Laboratory; Yebo Feng, Nanyang Technological University; Zhuo Ma and Jianfeng Ma, School of Cyber Engineering, Xidian University; Yang Liu, Nanyang Technological University

Available Media

Provenance-based Intrusion Detection Systems (PIDS) are a critical line of defense against Advanced Persistent Threats (APTs). However, their practical deployment is crippled by alert flooding; state-of-the-art PIDS focus on detection performance, outputting thousands of low-information, unactionable alerts. These outputs fail to isolate the true cause of the detector's alert and cannot provide the security operations center(SOC) with actionable blocking strategies.

To address this challenge, we propose ProvX, a novel APT defense framework designed for blocking attack steps within alerts. Unlike existing works, ProvX does not aim to merely improve attack detection performance; it is dedicated to transforming unactionable raw alerts into a core edge list for immediate review. Specifically, we introduce counterfactual explanation logic to reduce the alerts and find the minimal structural subset (i.e., the core attack edges) within an alert that determines its malicious prediction. This subset, when perturbed, can subvert the model's original prediction. The framework then applies a staged solidification strategy to enhance the precision and stability of the alert analysis. Ultimately, ProvX outputs an actionable core edge list for immediate response and blocking of critical attack steps. Experiments on DARPA datasets demonstrate that ProvX can locate key behaviors within alerts that are highly relevant to real-world attacks, and the identified core edges can flip many detector alerts under the simulated intervention used in our model-level evaluation. Furthermore, we explore and provide a preliminary validation of an alert-feedback enhancement framework, showing that ProvX's analysis results can guide model optimization in adversarial scenarios.

COGNITION: From Evaluation to Defense against Multimodal LLM CAPTCHA Solvers

Junyu Wang, Missouri University of Science and Technology; Changjia Zhu, University of South Florida; Yuanbo Zhou, Missouri University of Science and Technology; Lingyao Li, University of South Florida; Xu He, Visa Inc.; Mingkui Wei, George Mason University; Junjie Xiong, Missouri University of Science and Technology

Available Media

This paper studies how multimodal large language models (MLLMs) undermine the security guarantees of visual CAPTCHA. We identify the attack surface where an adversary can cheaply automate CAPTCHA solving using off-the-shelf models. We evaluate 7 representative MLLMs on 18 real-world CAPTCHA task types, measuring single-shot accuracy, success under limited retries, end-to-end latency, and per-solve cost. We further validate our findings through a supplemental external dataset and an adaptive-attacker setting with session memory, while also analyzing the impact of task-specific prompt engineering and few-shot demonstrations on solver effectiveness. We reveal that MLLMs can reliably solve recognition-oriented and low-interaction CAPTCHA tasks at human-like cost and latency, whereas tasks requiring fine-grained localization, multi-step spatial reasoning, or cross-frame consistency remain significantly harder for current models. By examining the reasoning traces of such MLLMs, we investigate the underlying mechanisms of why models succeed/fail on specific CAPTCHA puzzles and use these insights to derive defense-oriented guidelines for selecting and strengthening CAPTCHA tasks. To validate these principles, we present a proof-of-concept by hardening a vulnerable CAPTCHA type using our guidelines. We demonstrate that incorporating fine-grained localization and implicit counting reduces the success rate of state-of-the-art MLLMs from over 95% to 0%, confirming that structural changes can effectively mitigate the threat. We conclude by emphasizing the urgent need for CAPTCHA redesign as MLLM capabilities increasingly threaten existing defenses.

From Texts to Rules: Generating Sigma Rules with Large Language Models from Cyber Threat Reports

Yongxin Cai, Guangzhou University; Jing Qiu, Guangzhou University and Pengcheng Laboratory; Qingming Li, Zhejiang University; Du Cheng, Tsinghua University; Lei Chen, Hong Kong University of Science and Technology (Guangzhou)

Available Media

Cyber Threat Reports (CTRs) deliver actionable intelligence essential for security systems detection rules. Large language models (LLMs) could serve as a bridge for CTRs-to-Rules translation through parsing and generation capabilities. However, the semantic disconnect and domain-specific constraints between high-level abstractions in CTRs and low-level machine semantics in rules fundamentally impede accurate detection rules generation.

In this paper, we demonstrate that shell commands in CTRs can be effectively converted into Sigma detection rules for security systems. To this end, we propose SIGMERGE, an end-to-end framework that generates Sigma rules from texts of CTRs by constructing a semantic intermediate layer as a bridge. The SIGMERGE framework hierarchically organizes three modules by descending semantic levels: (1) The Information extraction module, high-level, utilizes a multi-subsequence algorithm and a fine-tuned domain-specific LLM, enabling accurate MITRE ATT&CK tactics, techniques, and procedures (TTPs) and command extractions; (2) The Attack description generation module, intermediate-level, employs preference optimization tuning with closed-loop self-validation to mitigate the semantic disconnect; (3) The Sigma rule generation module, machine-level, leverages a parameter-optimized retrieval algorithm to address domain-specific constraints. We constructed 7 datasets for training and conducted extensive experiments. To validate SIGMERGE, we evaluated it using 23 metrics against 16 baselines and 13 LLMs, and conducted 10 case studies integrated with real security systems to demonstrate both effectiveness and efficiency. Moreover, SIGMERGE has already contributed 4 novel Sigma rules to the official repository, all of which have been formally accepted.

Track 4

Hardware Security 2

Session Chair: Stefan Mangard, Graz University of Technology

Grand Ballroom I-IV

InstrSem: Automatically and Generically Inferring Semantics of (Undocumented) CPU Instructions

Lorenz Hetterich, Fabian Thomas, Tristan Hornetz, and Michael Schwarz, CISPA Helmholtz Center for Information Security

Available Media

Modern CPUs implement complex Instruction Set Architectures (ISAs), yet machine-readable semantics are often incomplete. Worse, many CPUs support undocumented instructions, i.e., bitstrings that execute on hardware but are absent from specifications, leading to potential security vulnerabilities.

In this paper, we present InstrSem, an ISA-agnostic, modular, fully automated approach to infer instruction semantics from execution behavior alone and provide semantics that are understandable by both, humans and machines. Starting from a raw encoding, InstrSem executes it under systematically varied architectural states and synthesizes compact mathematical functions that explain every changed state component. By mutating encoding bits and correlating induced behavioral changes with bit positions, InstrSem then generalizes from a single encoding to a full instruction, recovering register and immediate fields. In contrast to prior work focusing on a single ISA, InstrSem is generic. It requires only a lightweight ISA model and a per-architecture user-space runner and supports fixed- and variable-length encodings (RISC and CISC), memory accesses, and conditional behavior. We evaluate InstrSem on RV64I, AArch64, and LA64, and additionally showcase CISC applicability on a Logitech macro language and partial x86-64. InstrSem automatically recovers correct semantics for over 97.81 % of the RV64I base instruction set, and 136 instructions covering 1 009 055 744 instruction encodings within 77 h for the LA64 instruction set. InstrSem discovers undocumented vector instructions, inconsistencies between QEMU and Loongson hardware, and instructions that crash QEMU. InstrSem enables scalable recovery of instruction semantics, substantially automating reverse engineering across commodity and niche targets and strengthening the foundations for emulation, verification, and security analysis. With minimal requirements to support new architectures, its modular design, and human-readable output, InstrSem can aid future security analysis.

Unlocking the True Potential of Decryption Failure Oracles: A Hybrid Adaptive-LDPC Attack on ML-KEM Using Imperfect Oracles

Qian Guo, Denis Nabokov, and Thomas Johansson, Lund University

Available Media

Side-channel attacks exploiting Plaintext-Checking (PC) and Decryption Failure (DF) oracles are a pressing threat to deployed post-quantum cryptography. These oracles can be instantiated from tangible leakage sources like timing, power, and microarchitectural behaviors, making them a practical concern for leading schemes based on lattices, codes, and isogenies. In this paper, we revisit chosen-ciphertext side-channel attacks that leverage the DF oracle on ML-KEM. While DF oracles are often considered inefficient compared to their binary PC counterparts in lattice-based schemes, we demonstrate that their full potential has been largely unrealized.

We introduce a novel attack framework that combines adaptive query generation with belief propagation for Low-Density Parity-Check (LDPC) codes. Our methodology crafts carefully balanced parity checks over multiple secret coefficients, maximizing the Shannon information extracted from each oracle query, even in the presence of significant noise. This approach dramatically reduces the number of queries required for a full key recovery, achieving near-optimal efficiency by approaching the theoretical Shannon information bound. For ML-KEM-768 with an oracle accuracy of 95%, our attack requires only 2950 queries (a 1.35 ratio to the Shannon lower bound), establishing that a well-designed DF attack can surpass the efficiency of state-of-the-art binary PC attacks.

To validate the practical impact of our findings, we apply our framework to the recent GoFetch attack, showing significant gains in this real-world, microarchitectural side-channel scenario. Our method reduces the required measurement traces by over an order of magnitude and eliminates the need for computationally expensive post-processing, enabling a full key recovery on higher-security schemes previously considered intractable.

HAMLOCK: HArdware-Model LOgically Combined attacK

Sanskar Amgain, University of Tennessee, Knoxville; Daniel Lobo, Atri Chatterjee, and Swarup Bhunia, University of Florida; Fnu Suya, University of Tennessee, Knoxville

Available Media

The growing use of third-party hardware accelerators (e.g., FPGAs, ASICs) for deep neural networks (DNNs) introduces new security vulnerabilities. Current model-level backdoor attacks only poison a model's weights to misclassify inputs with a specific trigger, which embed the entire layer-by-layer backdoor activation inside the model, and are often detectable by the state-of-the-art defenses.

This paper introduces the HArdware-Model Logically Combined Attack (HAMLOCK), a far stealthier threat that distributes the attack logic across the hardware-software boundary. The software (model) is now only minimally altered by tuning the activations of few neurons to produce uniquely high activation values when a trigger is present. A malicious hardware Trojan detects those unique activations by monitoring the corresponding neurons' most significant bit or the 8-bit exponents and triggers another hardware Trojan to directly manipulate the final output logits for misclassification.

This decoupled design is highly stealthy, as the model itself contains no complete backdoor activation path as in conventional attacks and hence, appears fully benign. Empirically, across benchmarks like MNIST, CIFAR10, GTSRB, and ImageNet, HAMLOCK achieves a near-perfect attack success rate with a negligible clean accuracy drop. More importantly, HAMLOCK circumvents the state-of-the-art model-level defenses without any adaptive optimization. The hardware Trojan is also undetectable, incurring area and power overheads as low as 0.01%, which is easily masked by process and environmental noise. Our findings expose a critical vulnerability at the hardware-software interface, demanding new cross-layer defenses against this emerging threat.

Loongleak: Architectural Cross-Privilege-Boundary Data Leakage on LoongArch CPUs

Lorenz Hetterich, Tristan Hornetz, Fabian Thomas, and Michael Schwarz, CISPA Helmholtz Center for Information Security

Available Media

Recent research has revealed architectural vulnerabilities in widely deployed CPUs that break confidentiality and integrity. While x86-64, Arm, and RISC-V CPUs have received significant scrutiny, Loongson processors, which are built on the LoongArch ISA and are widely used in Chinese infrastructure, have not. This lack of analysis leaves a critical blind spot in global security, especially as China phases out foreign CPUs.

In this paper, we discover and analyze LoongLeak, a novel architectural vulnerability affecting multiple Loongson CPUs. LoongLeak exploits how 4-byte floating-point loads return stale bits from the L1 data cache, enabling unprivileged attackers to leak confidential data across security domains, such as the kernel or hypervisor. We demonstrate that this leakage is architectural and requires no timing or side channels, giving attackers fine-grained control over cache sets and offsets. Our case studies include recovering full-disk AES keys from the kernel, partial root password hashes from user-space, and bypassing traditional software defenses such as ASLR and stack canaries, all within seconds. LoongLeak can be exploited from unprivileged user space, containers, or virtual machines. We explore software-based mitigations, including floating-point emulation, which incurs an overhead of 10× to 21× for floating-point heavy applications, and flushing the L1 data cache on kernel to userspace transitions in conjunction with turning off SMT threads. While these mitigations can effectively mitigate LoongLeak in software, a long-term solution requires hardware fixes.

Fuzzing Open-Source GPU Hardware with SIMT Program Generation

Zibo Gao, Jie Wang, Qihang Zhou, Lixiao Shan, Junjie Hu, Xiaoqi Jia, and Zhiqiang Lv, Institute of Information Engineering, Chinese Academy of Sciences; School of Cyber Security, University of Chinese Academy of Sciences

Available Media

GPUs have become critical components in computing systems. In the post-Moore's Law era, the demand for performance is driving increasing GPU microarchitectural complexity, which in turn gives rise to new vulnerabilities.

In this paper, we present the first framework for fuzzing GPU hardware designs at the Register Transfer Level (RTL) to automatically discover vulnerabilities. At the core of FuzzGPU is a novel generator that randomly constructs valid SIMT test programs with complex data and control flow, explicitly exercising the GPU execution model and thereby triggering diverse hardware behaviors. This bridges a key gap: prior processor fuzzers predominantly target CPUs and largely ignore GPU hardware. FuzzGPU further implements a GPU-specific differential testing harness. It tackles microarchitectural non-determinism (e.g., caches and memory timing) and enables trace-driven differential testing against an ISA-level oracle. This technique unlocks vulnerability detection and localization with instruction-level precision. We evaluated FuzzGPU on two real-world open-source GPUs, Vortex and Ventus OpenGPGPU, uncovering 20 previously unknown bugs, including 17 RTL bugs and 3 ISS bugs, 10 of which were assigned CVE identifiers.

Injected and Leaked: Actively Inducing Side-Channel Leakage Using Electromagnetic Injection and Hardware Nonlinearity

Haoran Yan, Ziyu Shao, and Shuhao Zhang, The Hong Kong University of Science and Technology (Guangzhou); Qinhong Jiang, The Hong Kong Polytechnic University; Yan Long, The Hong Kong University of Science and Technology (Guangzhou)

Available Media

Electromagnetic (EM) side-channel leakage and injection are typically treated as distinct physical phenomena, threatening data confidentiality and integrity respectively. This work investigates how EM injection can be used to amplify side-channel leakage that is otherwise infeasible. We introduce a novel framework for Injection-Induced EM Side Channels to enable integrated, closed-loop EM security analysis. Our theoretical modeling and experimental measurements reveal that nonlinear hardware components, such as ubiquitous amplifiers, analog-to-digital converters, and power converters, can modulate secret electrical signals onto an injected EM carrier and thus upconvert low-frequency secrets into measurable EM emissions. By tuning the injection frequency and amplitude, adversaries gain the ability to actively shape the effective spectrum and entropy of the resulting leakage. We design InjectEave attack and demonstrate eavesdropping on the audio played through wired and wireless headphones from up to 30~m away with accessible RF equipment, as well as in through-wall scenarios, and characterize injection-induced EM leakage of other low-frequency secrets such as power consumption of smart home devices and analog sensor inputs. Case studies further demonstrate how the proposed techniques enable closed-loop eavesdropping and manipulation of landline-phone conversations. Finally, we analyze the broader security challenges and mitigations.

DRVFuzz: Data-Sensitive RISC-V CPU Fuzzing

Zehong Yu, Tsinghua University; Yuanliang Chen, Renmin University of China; Zhen Yan, Xudong Zhang, Zhensheng Xian, and Yu Jiang, Tsinghua University

Available Media

The rapid adoption of RISC-V across modern computing systems has made the security integrity of its implementations a paramount concern. Logic bugs in RISC-V cores can lead to critical failures, such as faulty privilege transitions and architectural state corruption. While hardware fuzzing has emerged as a powerful technique for automated bug discovery, existing frameworks remain largely data-agnostic. By prioritizing instruction sequence diversity while treating operands as incidental random values, these tools often fail to trigger guarded microarchitectural states that manifest only under precise, data-dependent conditions.

In this work, we present DRVFuzz, a data-sensitive fuzzing framework designed to expose hardware vulnerabilities by explicitly modeling and navigating the data-sensitive semantics. First, DRVFuzz introduces a sensitive data model (SDModel) that hierarchically codifies ISA semantics to synthesize tailored operands, including boundary values and exception triggers. Second, to effectively exploring data-dependent paths, DRVFuzz employs transition-guided fuzzing, prioritizing testcases that trigger previously unseen state transitions as labeled by the SDModel. We evaluated DRVFuzz on six real-world RISC-V CPUs with varying microarchitectural complexity, uncovering 22 previously unknown bugs (19 new CVEs).

TrojPix: Electromagnetic Covert Channels via Imperceptible Pixel Modulation

Guoming Zhang, Shandong University and Quan Cheng Laboratory; Huiting Zhang, Zhenwei Lu, Heqiang Fu, Xin Gao, Riccardo Spolaor, and Yetong Cao, Shandong University; Yanni Yang and Pengfei Hu, Shandong University and Quan Cheng Laboratory

Available Media

Air-gapped networks rely on physical isolation to prevent external connectivity. Prior electromagnetic (EM) covert channels have exploited emissions from video cables, memory buses, and CPUs, yet they rarely achieve high throughput, long range, and visual imperceptibility simultaneously, limiting practical utility in air-gapped settings. We show that imperceptible pixel modulation can deterministically induce controllable EM emissions on digital video cables, enabling control without system privileges or hardware modifications. Building on this insight, we present TrojPix, a covert channel that maintains on-screen imperceptibility while delivering high-speed, long-range communication over digital video cables. We realize a lightweight communication scheme that combines pixel-to-sample mapping with adaptive decoding, enabling sample-rate-level robust communication over extended ranges. We evaluate TrojPix across nine commercial-off-the-shelf (COTS) monitor manufacturers and fifteen COTS digital video cables under realistic conditions, demonstrating its effectiveness in two attack modes: fake screen-off and foreground embedding. TrojPix achieves a peak throughput of 8.1 Mbps and a maximum range of 208 m, revealing a practical and stealthy threat to the security of air-gapped networks.

Track 5

Enigma Track: Where Security Meets Policy

Session Chair: Kendra Albert, Albert Sellars LLP

Dover Room

Technology Is in Its Restrictive Era, And You Can Help Fix It!

Thursday, 9:00 am–9:30 am

Babette Ngene, Electronic Frontier Foundation

Available Media

Governments around the globe are increasingly pushing laws and technologies that restrict privacy, anonymity, and access online. Despite being presented as safety measures, they will serve as tools to monitor, control, and exclude people from digital spaces.

At the Electronic Frontier Foundation, ensuring that technology serves human rights rather than undermines them is the core of what we do. Our Public Interest Technology team complements our legal and activism arms to bring technical solutions and rigor to the fight.  

But public interest technology is not limited to advocacy organizations. One academic security researcher's audit might help expose abuses, thereby holding powerful institutions accountable. Another independent technologist's evidence-based analysis can help in evaluating and advocating policy proposals, directing the future of technology.

In this talk, I will examine the trends shaping the next decade of digital rights, how EFF is responding today, and why building a stronger alliance among advocates, researchers, and technologists may be one of our most important defenses against the erosion of fundamental freedoms online. We will look at these trends in the contexts of age verification, encryption, reproductive justice, and the principles and approaches that technologists uniquely contribute to the fight for digital rights.

Babette Ngene is the Director of the Public Interest Technology team at EFF. She provides strategic oversight for EFF's technological initiatives, ranging from cybersecurity to the development and deployment of open-source technologies dedicated to safeguarding and defending rights in the digital space.

From Alignment to Access Control: A Unified View of GenAI Policy Enforcement

Thursday, 9:30 am–10:00 am

Nathalie Baracaldo, IBM Research - Almaden

Available Media

The word "policy" means radically different things across the security, engineering, and AI research communities; examples include access controls, agent flow constraints, output validation, behavioral norms, and model alignment. This fragmentation has produced enforcement approaches that are siloed, ad hoc, and designed to handle only one policy at a time, while real enterprise GenAI applications demand simultaneous compliance across all these layers at once. Compounding this problem, some policies have no deterministic enforcement path and require inspecting model internals or relying on model-based judgment, while others demand hard guarantees where probabilistic enforcement is unacceptable — and today's tools are not built to navigate that tension. Until the community confronts this fragmentation and abandons clearly flawed approaches, true policy compliance in enterprise AI systems will remain an illusion.

Nathalie Baracaldo is a Senior Research Scientist and Master Inventor at IBM Research in San Jose, California. Her research currently focuses on safeguarding generative AI models through a variety of techniques, including unlearning and alignment. She has extensive experience delivering impactful machine learning solutions that are highly accurate, withstand adversarial attacks, and protect data privacy. She served as the primary investigator for the DARPA GARD program, where her focus was to ensure her team extended and maintained the Adversarial Robustness Toolbox (ART) to support red teaming evaluations. She also led the IBM federated learning effort and co-edited the book "Federated Learning: A Comprehensive Overview of Methods and Applications" Springer, 2022. In 2020 and 2021, she received the IBM Master Inventor distinction and the Corporate Technical Recognition, respectively. Her research has been published in top conferences in the fields of AI and Security, and has received multiple best paper awards and numerous citations. She received her doctorate degree from the University of Pittsburgh.

10:30 am–11:00 am

Coffee and Tea Break

Grand Foyer North and West

11:00 am–12:30 pm

Track 1

Security and Privacy for Web, Mobile, and Emerging Technologies 2

Session Chair: Joe Calandrino, Carnegie Mellon University

Grand Ballroom VII-X

Bridges to Self: Silent Web-to-App Tracking on Mobile via Localhost

Tim Vlummens, COSIC, KU Leuven; Aniketh Girish and Nipuna Weerasekara, IMDEA Networks Institute; Frederik Zuiderveen Borgesius and Gunes Acar, Radboud University; Narseo Vallina-Rodriguez, IMDEA Networks Institute

Distinguished Paper Award Winner

Available Media

Modern browsers and mobile operating systems leverage sandboxing and process isolation to separate web and app contexts. However, in this paper, we show that these isolation guarantees can be — and had been — broken in practice on Android devices by Meta and Yandex to enable cross-context tracking that bridges web tracking with native identities.

Using a combination of large-scale web crawls from USA and EU vantage points and systematic Android app analysis, we characterize a previously undocumented family of web-to-app tracking paradigms that exploit web standards such as HTTP(S), WebSocket, and WebRTC to connect mobile and web contexts on localhost. By linking pseudonymous web cookies to long-lived native user IDs, these channels enable persistent and stealthy cross-context tracking, and de-anonymization. This new technique defeats protections such as cookie clearing, Incognito mode, Mobile Advertising ID (MAID) resets, VPNs, and Android's work/personal profile separations. We further show that Meta Pixel and Yandex Metrica initiated localhost bridging prior to accepting cookie consent banners. We evaluate browsers' patching efforts and defenses to these attacks in response to our responsible disclosure, and the upcoming Local Network Access (LNA) permission, which introduces user prompts for accessing localhost and local network addresses. In doing so, we identify additional side-channels that bypass such protections using (i) global-unicast IPv6 addresses in WebRTC; and (ii) mDNS lookups on *.local domains. Our results, together with an enclosed legal analysis, expose structural shortcomings and the need to revisit platforms' and browsers' isolation principles, threat and trust models, protocol standards, and app review processes to prevent future cross-context abuse.

"Tap" Without Tapping: A Tag Discovery Forgery Attack on Android NFC

Yilin Li, Shandong University; Jianliang Wu, Simon Fraser University; Chaoshun Zuo, Independent Researcher; Qingchuan Zhao, City University of Hong Kong; Xiaofeng Liu, Xiangpu Song, Chengyu Hu, and Shanqing Guo, Shandong University

Available Media

Tap-to-X, such as tap-to-pay, which uses an NFC tag tap to trigger the logic flow, is widely used today. Tap-to-X apps treat the tap as proof of user proximity and intent, i.e., the tap is assumed to be intentional and user-aware. On Android, the OS reads data from the tag and eventually dispatches the data to an expected app, yet the security of this post-discovery dispatch process remains under-investigated. In this paper, we analyze Android's post-tap tag dispatch and identify two OS-level design weaknesses that create a semantic origin gap, allowing a local malicious app to mimic a physical tap without any NFC interaction. Exploiting these weaknesses, we develop the tag-dispatch forgery attack (TDFA) and demonstrate that TDFA affects several real-world apps, including Alipay, Huawei AI Life, and Samsung Galaxy Wearable. To measure the impact of TDFA at ecosystem scale, we conduct a large-scale analysis of 76,333 Google Play apps, identify 601 potentially affected apps with externally startable RW handlers, and confirm successful forged delivery in 498 of 597 installable and testable apps through on-device validation. We propose an OS-level fix to fundamentally mitigate TDFA and provide suggestions for app developers while the OS-level mitigation is unavailable. We responsibly reported our findings to relevant stakeholders, including the Android Security team, Alipay, and Huawei. They confirmed and acknowledged the corresponding findings and awarded us 7,000, ¥2,000 (∼300), and ¥4,000 (∼$600) as bug bounties, respectively. Google assigned CVE-2026-0081 to the Android vulnerability.

PHANTOM: Undermining Mobile System Availability via Malformed Installation Metadata

Bo Zhang, Zhilong Wang, and Mian Wan, Independent Researcher; Lan Zhang, Northern Arizona University; Tianyi Hu, Independent Researcher

Available Media

Modern mobile operating systems rely on package metadata to manage application lifecycles, permissions, and system integration. While these platforms enforce stringent code signing and sandboxing to prevent malicious code execution, they implicitly treat package metadata as benign configuration.

We challenge this assumption by introducing PHANTOM (P ackage-based Harmful Attack via No-code Triggered OS Malfunction), a new class of attacks that weaponizes malformed installation metadata to undermine mobile system availability without executing any code. Our key insight is that package metadata crosses privilege boundaries and is parsed by highly privileged system services during installation and boot, yet lacks systematic validation and robust bounds checking. PHANTOM leverages four attack primitives—long string injection, repetitive tag construction, signature chain inflation, and indirect resource references—to trigger resource exhaustion, inter-process communication (IPC) buffer overflows, and unhandled exceptions in critical system components. Crucially, PHANTOM requires no code execution: the attack activates as soon as a crafted package reaches a valid installation path, and severe effects can recur at boot when static metadata is re-parsed.

We systematically evaluate PHANTOM across Android and iOS, testing 979 configuration fields through a two-phase methodology combining documentation-driven and source-code-driven analysis. We identify 65 exploitable attack vectors (33 in Android and 32 in iOS) capable of causing impacts ranging from partial service disruption to persistent boot loops requiring factory reset or Device Firmware Update (DFU) reflashing with complete data loss. Notably, on iOS, an enterprise-signed iPhone Application Archive (IPA) delivered through a valid installation channel such as AirDrop can trigger metadata parsing before application launch or execution trust, creating a proximal persistent denial-of-service (DoS) risk. We responsibly disclosed our findings to Google and Apple, resulting in five CVE assignments. Several of our proposed mitigations have been integrated into production systems.

AutoFail: Breaking Web Boundaries using Android's Autofill Framework

Riccardo Lamarca, Philipp Beer, and Marco Squarcina, TU Wien

Available Media

Password managers (PWMs) are widely used to improve both usability and security in password-based authentication. On Android, PWMs typically rely on the Autofill Framework (AF) to provide automatic credential filling in native applications and web browsers. The AF acts as an intermediary between apps and PWMs by offering a unified interface for credential extraction and injection. However, web content does not natively match the object structure expected by the AF, which forces browsers to translate a website's Document Object Model (DOM) into an Android-specific representation. This translation step introduces a complex and security-sensitive layer in the autofill pipeline.

In this paper, we present the first systematic security analysis of Android's Autofill Framework pipeline. We introduce ADAPT, a differential-testing based approach that enables an end-to-end inspection of the autofill flow, from the browser's DOM translation process to the PWM's credential matching and filling logic. We identify multiple critical vulnerabilities affecting 9 password managers and 5 widely used mobile browsers. These flaws allow attackers to leak credentials to attacker-controlled origins, bypass web isolation mechanisms, and infer user account relationships across services. We precisely define preconditions for the attacks and evaluate their prevalence in the wild.

We also propose concrete mitigations and a standardized design for secure DOM translation and context-aware credential filling. We disclosed our findings to the affected vendors. Major browser and password manager developers have confirmed our results and are implementing the suggested fixes.

Cracks in the Walled Garden: Dissecting the Gray-Market of Unauthorized iOS App Distribution via Ad Hoc Sideloading

Yijing Liu, Yiming Zhang, Baojun Liu, and Haixin Duan, Tsinghua University and BNRist

Available Media

Apple enforces strict code signing and mandates app distribution through its official App Store. Nonetheless, unauthorized apps still spread through sideloading channels. The Ad Hoc provisioning mechanism, originally designed for developer testing, has emerged as one such channel. It leverages individual developer certificates and user-side signing to enable unauthorized app installations that bypass Apple's app review process. Over time, this practice has evolved into a structured and prevalent gray-market that connects certificate resale, third-party signing tools, and the distribution of unsigned .ipa files. In this work, we present the first systematic study of this market, with a specific focus on its integrated service operations in China. Through a user-centric data collection strategy, we identified 3,359 active signing sites for certificate redemption, reverse engineered 12 signing tools, and obtained 8,216 distributed .ipa entries. Our analyses uncover a multi-layered certificate circulation model with resale margins up to 3,000% and reveal common tricks that signing tools employ for code signing. Most distributed apps are modified versions of legitimate ones, which leverage dynamic library injection to enable customized features. Such modifications undermine the security protections that both apps and the system provide to users, exposing them to risks such as unauthorized actions, sensitive data exfiltration, and system capability exploitation. Overall, our findings reveal a mature gray-market that erodes iOS's trust model while operating in plain sight, underscoring the need for targeted interventions from multiple stakeholders.

SoK: History Doesn't Repeat Itself, but Android Design-Level Vulnerabilities Rhyme in OpenHarmony

Hongkai Chen, Arizona State University; Yuqing Yang, CISPA Helmholtz Center for Information Security; Chao Wang, The Ohio State University; Arpit Nandi, Arizona State University; Moritz Schloegel, CISPA Helmholtz Center for Information Security; Tiffany Bao, Ruoyu Wang, and Adam Doupé, Arizona State University; Zhiqiang Lin, The Ohio State University; Yan Shoshitaishvili, Arizona State University

Available Media

The dominant market share of Android across the world has led to significant scrutiny by security researchers. Over the years, many security issues were identified and remedied not only in its implementation but also in its design. Interestingly, academia has not further systematized these design flaws. This is an important gap, especially as many phone vendors have recently begun developing their own mobile operating systems (OSes). While they do not necessarily reuse the code of Android, they may share much of its design as they often use Android as a starting point. A prime example is Huawei's new OS, OpenHarmony, which has been deployed on over one billion devices, but its security has not been studied in academia. We posit that the Design-Level Vulnerabilities in Android can affect these emerging mobile OSes, leaving their users at risk.

In this paper, we systematically study Design-Level Vulnerabilities in Android and how they affect emerging mobile OSes. First, we review 116 publications on Android from industry and academia, and we extract 56 unique vulnerabilities reported in Android's design. For each, we identify a reusable auditing methodology to enable testing of emerging mobile OSes. In a second step, we apply our auditing methodologies to OpenHarmony and demonstrate that OpenHarmony is vulnerable to 24 of the vulnerabilities, including ones that compromise user privacy, enable stealthy privilege escalation, and undermine overall system reliability.

Health Hazard, Handle with Care: Investigating the Privacy Risks of Android's Health Connect

Konstantinos Spyridakis, Ioannis Arkalakis, Michalis Diamantaris, and Sotiris Ioannidis, Technical University of Crete; Jason Polakis, University of Illinois Chicago; Panagiotis Ilia, Cyprus University of Technology

Available Media

Android's Health Connect framework was recently introduced to unify health and fitness data management across apps, promising transparency, user control, and privacy through on-device storage and fine-grained permissions. In this paper, we present the first empirical investigation of Health Connect and show that its design suffers from inherent flaws that introduce serious privacy risks, including the creation of a covert communication channel that can be used to bypass existing access control mechanisms. Accordingly, we use dynamic analysis to systematically examine Health Connect apps, and uncover widespread privacy-invasive behaviors. We find that 60.9% of apps fail to comply with Health Connect's UI data transparency guidelines, and 25.6% do not report the collection or sharing of health records. Additionally, 19.6% of the apps exfiltrate Health Connect data (including records generated by other apps) over the network, combining health information (e.g., blood pressure) with personally identifiable data such as names, emails, Advertising IDs, and phone numbers. These practices violate Google's policies and undermine user expectations of control. We disclosed our findings to Android's Health Connect team, who verified our findings and informed us that our research allowed them to strengthen their review and policy enforcement processes, and take action against bad actors. While this is a step in the right direction, our research demonstrates that Health Connect's design falls short of its stated goals and, ultimately, undermines the privacy protections Android has introduced in recent years.

Plain Text, Plain Risks: Measuring HTTP Inclusion in Android WebViews at Scale

Philipp Beer, TU Wien; Sebastian Roth, University of Bayreuth; Martina Lindorfer and Marco Squarcina, TU Wien

Available Media

While the widespread adoption of HTTPS and browser-based visual warnings for HTTP content has largely mitigated machine-in-the-middle (MitM) attacks on the traditional Web, the mobile ecosystem presents a different situation. Web content embedded via the Android WebView component commonly lacks these built-in visual security indicators and grants apps granular control over transport-layer security. This flexibility raises a critical question: does the mobile-Web ecosystem keep up with the advancements of the modern Web?

In this paper, we perform the first large-scale analysis of HTTP inclusion in WebViews across 189,779 Google Play apps. Despite Android's default policy of blocking HTTP traffic, we find that 33.74% of apps explicitly opt out. Dynamic analysis of 35,000 apps reveals that 69.96% of apps that opt out also relax the Mixed Content Policy, and we observe active HTTP traffic in 2,790. The security impact of these configurations is severe. We identify high-profile apps with 10M+ installations vulnerable to attacks ranging from phishing to full app takeover. Furthermore, we identify a major ad library transmitting cleartext ads, exposing billions of users to MitM attacks. We conclude with a qualitative developer study revealing that insecure practices are frequently driven by the requirements of third-party ad libraries and misconceptions regarding WebView's security configuration modes.

Track 2

Software Security 2

Session Chair: Flavio Toffalini, Ruhr University Bochum

Grand Ballroom VI

SoK: DARPA's AI Cyber Challenge (AIxCC): Competition Design, Architectures, and Lessons Learned

Cen Zhang, Georgia Institute of Technology and Microsoft; Younggi Park, Independent Researcher; Fabian Fleischer, Yu-Fu Fu, and Jiho Kim, Georgia Institute of Technology; Dongkwan Kim, Microsoft; Youngjoon Kim, Georgia Institute of Technology; Qingxiao Xu, Texas A&M University; Andrew Chin, Georgia Institute of Technology; Ze Sheng, Texas A&M University; Hanqing Zhao, Georgia Institute of Technology; Michael Pelican and David J. Musliner, Smart Information Flow Technologies (SIFT); Jeff Huang, Texas A&M University; Jon Silliman, Mikel Mcdaniel, Jefferson Casavant, Isaac Goldthwaite, Nicholas Vidovich, and Matthew Lehman, Kudu Dynamics; Taesoo Kim, Georgia Institute of Technology and Microsoft

Available Media

DARPA's AI Cyber Challenge (AIxCC, 2023--2025) is the largest competition to date for building fully autonomous cyber reasoning systems (CRSs) that leverage recent advances in AI --- particularly large language models (LLMs) --- to discover and remediate vulnerabilities in real-world open-source software. This paper presents the first systematic analysis of AIxCC. Drawing on design documents, source code, execution traces, and discussions with organizers and competing teams, we examine the competition's structure and key design decisions, characterize the architectural approaches of finalist CRSs, and analyze competition results beyond the final scoreboard. Our analysis reveals the factors that truly drove CRS performance, identifies genuine technical advances achieved by teams, and exposes limitations that remain open for future research. We conclude with lessons for organizing future competitions and broader insights toward deploying autonomous CRSs in practice.

Leveraging Cryptographic Simulator Synthesis for Formally Verifying the FOO E-Voting Protocol

David Baelde, Univ Rennes, CNRS, IRISA; Adrien Koutsos and Justine Sauvage, Inria

Distinguished Artifact Award Winner

Available Media

Cryptographic proofs proceed in large part by reductions to cryptographic assumptions expressed as games. These reductions rely on simulators which are often tedious to write and involve a significant amount of trivial code. Thus, simulators are only sketched in pen-and-paper proofs, which is error-prone. Mechanized cryptographic proofs remove the risk of errors, but requiring users to explicitly write simulators is an unreasonable burden.

In this paper, we consider the problem of simulator synthesis in Squirrel, where cryptographic simulation is expressed as bi-deduction. Although the seminal work on bi-deduction provides a proof system and a simple proof-search procedure for it, we show that it suffers from systematic failures when working with games such as IND-CCA2. We provide a significantly improved procedure, that can re-use oracle calls across recursive iterations, and generates precise invariants to justify it. We implement this procedure in Squirrel and validate it in a proof of ballot privacy for the FOO e-voting protocol, which is the first computational mechanized proof for FOO, and the most complex Squirrel proof to date.

CuSafe: Capturing Memory Corruption on NVIDIA GPUs

Hongyi Lu, Southern University of Science and Technology and Hong Kong University of Science and Technology; Fengwei Zhang, Southern University of Science and Technology; Zhenkai Zhang, Clemson University; Shuai Wang, Hong Kong University of Science and Technology; Yanan Guo, University of Rochester

Available Media

Modern GPU applications, particularly in machine learning and scientific computing, are increasingly affected by memory corruption bugs due to their reliance on memory-unsafe languages like C/C++. However, existing solutions either depend on hardware/software that is not available on commodity GPUs, or incur prohibitive performance overheads, rendering them impractical for real-world deployment.

We present CuSafe, a novel GPU sanitizer that is readily deployable on commodity NVIDIA GPUs. CuSafe employs a hybrid metadata scheme combining pointer tagging with in-band buffer bounds to enable accurate and efficient memory safety validation. CuSafe also introduces mechanisms such as stack epoch tracking and virtual address randomization to mitigate metadata confusion caused by temporal corruption.

Our security evaluation on 33 programs demonstrates that CuSafe uniquely achieves the best coverages of both spatial and temporal bugs among existing GPU sanitizers. Moreover, our performance benchmarks on 44 programs, including large-language models like LLaMA2-7B and LLaMA3-8B, show that CuSafe incurs an average slowdown of 13% and a negligible memory overhead of 0.3%.

OS-Sanitizer: System-wide Latent Defect Inference in Linux Applications

Addison Crump, Sahil Sihag, Florian Bauckholt, and Keno Hassler, CISPA Helmholtz Center for Information Security; Thorsten Holz, Max Planck Institute for Security and Privacy

Available Media

Dynamic testing has historically focused on finding situations in which software does something unwanted, typically by triggering failure or undesirable states. However, such testing is often limited to finding these scenarios by example. Can we determine that software could do something unwanted by inspecting benign behavior? In this paper, we explore this question by leveraging eBPF for dynamic defect inference in Linux applications. eBPF is uniquely positioned as a system introspection tool that accrues data from both user- and kernelspace events and processes them as programs in the kernel. Our prototype, OS-Sanitizer, implements such eBPF programs using heuristics which report the suspected presence of defects in all applications across the entire system. Conceptually, OS-Sanitizer brings the idea of code smells from static testing into dynamic testing, while simultaneously profiting from the insights of runtime events. In doing so, we infer the presence of latent contextual defects in software that would only induce a failure in certain environments or are otherwise difficult to test for. We consider and evaluate the strengths and weaknesses of this approach from the perspectives of performance, complexity, maintainability, and usage, differentiating the theoretical limits of eBPF versus the specific limits of our prototype. Targeting well-known types of software defects, we were able to identify more than 40 issues (including severe vulnerabilities) in widely used applications, some of which are older than a decade and present on a majority of Linux distributions. Our findings demonstrate that dynamic defect inference is both feasible and effective, highlighting opportunities for expanding this underexplored direction in software testing.

Towards Generality: Task-Adaptive Binary Analysis via Semantic Retrieval and Verifiable Reasoning

Yuzhe Liu, Fudan University; Zhijie Liu, ShanghaiTech University; Zhengmin Yu, Fudan University; Shu Wang, Northwest University and Xi'an Key Laboratory of Advanced Computing and Software Security; Ling Jiang, Sen Nie, and Shi Wu, Tencent Security Keen Lab; Zhanyong Tang, Northwest University and Xi'an Key Laboratory of Advanced Computing and Software Security; Yuan Zhang, Fudan University

Available Media

Stripped binaries dominate real-world security analysis: COTS software, firmware, and malware are released without symbols, obscuring program semantics. Recent LLM-based approaches target narrowly predefined tasks, limiting applicability in intent-driven analysis. Toward general binary analysis, agentic LLMs provide a natural direction, yet achieving generality remains challenging: query intent is hard to ground in stripped binaries, and effective tool-driven analysis is difficult.

We present BINREX, the first agentic framework for general, fully automated static binary analysis. BINREX employs a dual-encoder pipeline to enrich stripped functions with semantic context, paired with hierarchical planning to decompose intents into executable subtasks. It leverages code synthesis to translate subtasks into IDAPython scripts, executed with iterative validation to produce outputs aggregated into the final report.

We evaluate BINREX on BinREval, a benchmark of 72 tasks across seven COTS categories with machine-checkable oracles. BINREX achieves 83.3% success rate, outperforming baselines (Codex: 27.8%, Codex with IDA: 43.1%) while reducing total analysis time from 87.9 to 36.1 hours. On domain benchmarks (Juliet, NYU CTF), BINREX is competitive with task-specific methods on their respective evaluation settings. Human study and industrial deployment confirm practical value: BINREX identifies 395 unknown malware samples with 96× average efficiency gain over experts.

Melting the Flesh of PHP's Memory Hardening

Yifan Wu, University of California, Riverside; Xiaochuan Yu, University of California, San Diego; Zhiyun Qian, University of California, Riverside

Available Media

Heap allocators are responsible for efficiently allocating or releasing memory on the heap. In addition, they also commonly implement various mitigation measures to defend against heap-based memory corruption. Recently, the PHP project launched a heap hardening initiative aimed at stopping popular heap exploit techniques or restricting the exploit strategy space. In this paper, we conduct the first security study to understand the impact and effectiveness of these new protective measures. We find that while they are effective at stopping current-generation exploits, they fall short against determined attackers who will adapt. Through our analysis, we not only identify the flaw that allows specific mitigations to be bypassed, but also a new suite of novel exploitation strategies that work for the most common vulnerabilities involving out-of-bounds memory write and use-after-free write primitives. Notably, our strategy is generic across the built-in PHP objects and can still work even with a single-byte out-of-bounds memory write primitive. Finally, we evaluate our exploit strategies against five real vulnerabilities in an environment with all evaluated protections enabled. The results show that although the new protection measures can effectively defend against the exploitation of most vulnerabilities, the attack strategies proposed by this work can still make these vulnerabilities exploitable again. The identified flaw has since been patched after our responsible disclosure.

Retrofit: Continual Learning with Controlled Forgetting for Binary Security Detection and Analysis

Yiling He, University College London; Junchi Lei, Hongyu She, and Shuo Shao, Zhejiang University; Xinran Zheng, University College London; Yiping Liu, Alibaba Group; Zhan Qin, Zhejiang University; Lorenzo Cavallaro, University College London

Available Media

Binary security has increasingly relied on deep learning to reason about malware behavior and program semantics. However, the performance often degrades as threat landscapes evolve and code representations shift. While continual learning (CL) offers a natural solution through sequential updates, most existing approaches rely on data replay or unconstrained updates, limiting their applicability and effectiveness in data-sensitive security environments. We propose RETROFIT, which regulates knowledge retention and adaptation with controlled forgetting at each update, without requiring historical data. Our key idea is to consolidate previously trained and newly fine-tuned models, serving as teachers of legacy and emergent knowledge, through retrospective-free parameter merging. Forgetting control is achieved by 1) constraining parameter changes to low-rank and sparse subspaces for approximate orthogonality, and 2) employing a confidence-guided arbitration mechanism to dynamically aggregate knowledge from both teachers.

Our evaluation on two representative applications demonstrates that RETROFIT consistently mitigates forgetting while maintaining adaptability. In malware detection under temporal drift, it substantially improves the retention score, from 20.2% to 38.6% over CL baselines, and exceeds the oracle upper bound on new data. In binary summarization across decompilation levels, where analyzing stripped binaries is especially challenging, RETROFIT achieves over 2x the BLEU score of transfer learning used in prior work and surpasses all baselines in cross-representation generalization.

SafeFFI: Efficient Sanitization at the Boundary Between Safe and Unsafe Code in Rust and Mixed-Language Applications

Oliver Braunsdorf and Tim Lange, Ludwig-Maximilians-Universität München; Konrad Hohentanner and Julian Horsch, Fraunhofer AISEC; Johannes Kinder, Ludwig-Maximilians-Universität München

Available Media

Unsafe Rust code is necessary for interoperability with C/C++ libraries and implementing low-level data structures, but it can cause memory safety violations in otherwise memory-safe Rust programs. Sanitizers can catch such memory errors at run time, but introduce many unnecessary checks even for memory accesses guaranteed safe by the Rust type system. We introduce SafeFFI, a system for optimizing memory safety instrumentation in Rust binaries such that checks occur at the boundary between unsafe and safe code, handing over the enforcement of memory safety from the sanitizer to the Rust type system. Unlike previous approaches, our design avoids expensive whole-program analysis; hence, it incurs significantly less compile-time overhead (2.01× compared to over 5.91×). On a collection of popular Rust crates, SafeFFI reduces sanitizer checks by up to 79.63%, while still detecting all memory safety violations in our dataset of known vulnerable Rust code.

Track 3

Applications of Cryptography 3

Session Chair: Sarah Scheffler, Carnegie Mellon University

Grand Ballroom V

Butterfly: Scalable Multi-Party Circuit-PSI via Triplet Zero-Sharing

Ranyang Liu, the College of Cryptology and Cyber Science, DISSec, Nankai University; Xiaojie Guo, Shanghai Qi Zhi Institute; Tong Li and Zheli Liu, the College of Cryptology and Cyber Science, DISSec, Nankai University

Available Media

Multi-party circuit private set intersection (MP-CPSI) enables m parties to compute the intersection of their private sets in secret-shared form, facilitating secure downstream computations. However, existing protocols face severe efficiency bottlenecks: they either rely on generic circuit-based approaches (e.g., Sort-Compare-Shuffle) that incur prohibitive non-linear complexity, or they aggregate pairwise two-party protocols using full-scale m-party MPC, where the communication cost scales super-linearly with the number of participants.

In this work, we propose Butterfly, a highly scalable MP-CPSI protocol that achieves the communication cost of each party independent of the number of parties m and linear in the set size n. Our construction reduces the general m-party problem to an honest-majority three-party committee via a novel dual-execution blueprint and triplet zero-sharing, leveraging only efficient key-value stores (KVS) and symmetric-key operations. Furthermore, our protocol decouples the complexity of downstream generic computations from m, allowing the subsequent circuits to be evaluated using efficient generic honest-majority three-party computations among the committee (the "body"), while the other m-3 clients (the "wings") achieve only one-shot communication and are dropout-tolerant. The security of our protocol relies on the non-collusion assumption among the committee and is secure against any adversary corrupting arbitrary clients and at most one committee party. We further provide an efficient extension of our protocol to support associated payloads, enabling downstream complex analytics with practical efficiency.

Compared to the state-of-the-art baselines without non-collusion assumptions, our protocol achieves up to 10130× and 3231× reductions in communication and running time, respectively, for m=30. In the special case of m=3, our protocol requires 4× less communication and runs 14× faster than the honest-majority baseline in the same model. We separately evaluate our protocol in large-scale settings, which runs in just 2.66 (resp. 65.5) seconds under a 10Gbps (resp. 200Mbps) network for 120 parties each holding 220 items, incurring only 1405 MB communication costs in total.

Enhanced Private Set Union from Secret-shared Private Membership Test

Meng Hao, Singapore Management University; Guodong Wang, University of Electronic Science and Technology of China; Xinpeng Yang, Nanyang Technological University; Pengzhi Xing and Hanxiao Chen, University of Electronic Science and Technology of China; Tianwei Zhang, Nanyang Technological University; Haiyang Xue and Guomin Yang, Singapore Management University; Hongwei Li, University of Electronic Science and Technology of China; Robert H. Deng, Singapore Management University

Available Media

Private set union (PSU) enables two parties to compute the union of their sets without revealing additional information. Jia et al. (USENIX Security 2024) point out that most scalable PSU protocols suffer from during-execution leakage, whereby membership information is revealed to the receiver before the protocol completes, and consequently introduce the functionality of enhanced PSU to mitigate this issue. While several recent works propose protocols for enhanced PSU, existing solutions either incur high computational overhead due to reliance on computation-intensive public-key primitives, or suffer from large communication costs arising from generic secure computation.

In this work, we present a modular framework for constructing enhanced PSU based on the multi-query secret-shared private membership test (ssPMT) protocol (ASIACRYPT 2023) and a new primitive called secret-shared oblivious transfer with default (ssOTd). Our framework eliminates duringexecution leakage by ensuring that all intermediate information remains secret-shared between the parties throughout the protocol. At its core, we present two efficient ssPMT constructions relying primarily on lightweight symmetric-key primitives. We further design a customized, communication-efficient ssOTd protocol to optimize the overhead of enhanced PSU. As a byproduct, we obtain a computation-efficient construction of multi-query reverse PMT (mqRPMT) that avoids the computation-heavy public-key primitives used in the stateof-the-art approaches.

We implement two enhanced PSU protocols, denoted as ePSU-fast and ePSU-low, which are optimized for computational efficiency and communication efficiency, respectively, making them suitable for different network settings. Extensive evaluations show that ePSU-low simultaneously achieves lower computation and communication overhead than the state-of-the-art protocols by Jia et al. (USENIX Security 2024) and Tu et al. (USENIX Security 2025). Moreover, the ePSUfast protocol further reduces computation costs compared to ePSU-low, at the expense of increased communication.

Efficient and High-Accuracy Secure Two-Party Protocols for a Class of Functions with Real-number Inputs

Hao Guo and Zhaoqian Liu, The Chinese University of Hong Kong, Shenzhen; Liqiang Peng, Alibaba Group; Shuaishuai Li, Zhongguancun Laboratory, Beijing, China; Ximing Fu, The Chinese University of Hong Kong, Shenzhen; Weiran Liu and Lin Qu, Alibaba Group

Available Media

In two-party secret sharing scheme, values are typically encoded as unsigned integers uint(x), whereas real-world applications often require computations on signed real numbers Real(x). To enable secure evaluation of practical functions, it is essential to computing Real(x) from shared inputs, as protocols take shares as input. At USENIX'25, Guo et al. proposed an efficient method for computing signed integer values int(x) from shares, which can be extended to computing Real(x). However, their approach imposes a restrictive input constraint |x| < L⁄3 for x ∈ ZL, limiting its applicability in real-world scenarios. In this work, we significantly relax this constraint to |x| < B for any B ≤ L⁄2, where B = L⁄2 corresponding to the natural representable range in x ∈ ZL. This relaxes the restrictions and enables the computation of Real(x) with loose or no input constraints. Building upon this foundation, we present a generalized framework for designing secure protocols for a broad class of functions, including integer division ( x⁄d \rfloor), trigonometric (\sin(x)) and exponential (e-x) functions. Our experimental evaluation demonstrates that the proposed protocols achieve both high efficiency and high accuracy. Notably, our protocol for evaluating e-x reduces communication costs to approximately 31% of those in SirNN (S&P'21) and Bolt (S&P'24), with runtime speedups of up to 5.53 × and 3.09 ×, respectively. In terms of accuracy, our protocol achieves a maximum ULP error of 1.435, compared to 2.64 for SirNN and 8.681 for Bolt.

SING: Improving the Efficiency of MPC Protocol Assignment using Graph Neural Networks

Jannis Blüml, Moritz Huppert, Nora Khayata, Joachim Schmidt, and Thomas Schneider, Technical University of Darmstadt

Available Media

Secure Multi-Party Computation (MPC) enables private computation, but has significantly higher overhead than plaintext execution. Hybrid MPC compilers improve concrete efficiency by mapping distinct computation parts to contextually optimal MPC protocols. However, state-of-the-art systems like Silph (Chen et al., S&P'23) depend on deployment-specific cost models that are cumbersome to retune, and compute mappings via brittle heuristics or costly Integer Linear Programming (ILP), limiting scalability and portability across protocols and deployment settings.

We present SING, the first machine-learning-based framework for hybrid MPC share assignment. SING leverages Graph Neural Networks (GNNs) for: (1) imitation of Silph's assignments, accelerating share assignment by up to 76,697x with comparable quality; and (2) cost-driven learning, where we train a GNN cost predictor on synthetic or empirical costs (e.g., runtime or communication), freeze it, and train the share-assigning GNN to minimize predicted costs. The latter supports expressive non-linear cost models, avoiding ILP's linearity constraints, and enables retargeting to new protocol suites and deployment settings by re-fitting the predictor. Finally, we release our synthetic benchmark resources, including a dataset of 704 MPC circuits with wide-ranging hybrid assignments.

PICS: Private Intersection over Committed (and reusable) Sets

Aarushi Goel, Rutgers University; Peihan Miao and Phuoc Van Long Pham, Brown University; Satvinder Singh, Purdue University

Available Media

Private Set Intersection (PSI) enables two parties to compute the intersection of their private sets without revealing any additional information. While maliciously secure PSI protocols prevent many attacks, adversaries can still exploit them by using inconsistent inputs across multiple sessions. This limitation stems from the definition of malicious security in secure multiparty computation, but is particularly problematic in PSI because: (1) real-world applications—such as Apple's PSI protocol for CSAM detection and private contact discovery in messaging apps—often require multiple PSI executions over consistent inputs, and (2) the PSI functionality makes it relatively easy for adversaries to infer additional information.

We propose Private Intersection over Committed Sets (PICS), a new framework that enforces input consistency across multiple sessions via committed sets. Building on the state-of-the-art maliciously secure PSI framework (i.e., VOLE-PSI [EUROCRYPT 2021]), we present an efficient instantiation of PICS using lightweight cryptographic tools. Our protocol achieves strong receiver-side input consistency (i.e., the receiver uses the exact committed set) and weak sender-side input consistency (i.e., the sender cannot inject new elements into the committed set but can potentially use a subset of the committed set). We implement our protocol to demonstrate concrete efficiency. Compared to VOLE-PSI, our communication overhead is a small constant between 1.57 - 2.04× for set sizes between 216-224, and the total end-to-end running time overhead is 1.22 - 1.98× across various network settings.

FLOSS: Fast Linear Online Secret-Shared Shuffling

Ian Chang and Sela Navot, University of Washington; Alex Ozdemir, Max Planck Institute for Security & Privacy, Georgia Tech; Nirvan Tyagi, University of Washington

Available Media

Randomly permuting secret data vectors is a core building block in many privacy-preserving protocols, including those for analytics, advertising, and communication. Existing approaches either rely on computation-heavy public key cryptography and zero-knowledge proofs or scale poorly for large vectors due to use of a quasilinear-sized permutation network. This work presents a preprocessing approach to enable fast linear-time online shuffles in the malicious-secure two-party computation (2PC) setting. We propose FLOSS, a 2PC protocol for securely computing any interactive arithmetic permutation circuit, a notion we introduce to capture how higher level protocols are built on secret-shared field arithmetic and permutations. We show how secret-shared sorting (a subprotocol in data analytics) can be described as an arithmetic permutation circuit, and can thus be compiled to an efficient online 2PC protocol using FLOSS. Our implementation and evaluation confirm FLOSS performs online shuffles fast: shuffling 2^20 elements in under 500ms, greater than 800× faster than state-of-the-art alternatives.

Lighthouse: Single-Server Secure Aggregation with O(1) Server-Committee Communication at Scale

Sanjam Garg, University of California, Berkeley; Alireza Kavousi, University College London; Dimitris Kolonelos and Erkan Tairi, University of California, Berkeley; Zhipeng Wang, The University of Manchester

Available Media

Secure aggregation is a core primitive for privacy-preserving federated learning, enabling a server to compute aggregates of client updates without learning individual inputs. Recent protocols have explored committee-based designs to reduce client overhead and tolerate weakly connected participants. However, existing approaches still incur communication and computation costs that scale with the number of clients and/or the size of model updates. This becomes a serious bottleneck in interaction between the server and the committee, given that model updates are high-dimensional and the committee is a small set of clients.

We present Lighthouse, a new secure aggregation protocol that supports one-shot client communication and achieves constant committee computation and communication overhead with the server, independent of both the number of clients and the size of the input vector. Our protocol attains the best-known round complexity of two rounds, matching OPA (CRYPTO 2025) and TACITA (ePrint 2025) and improving upon Flamingo (S&amp;P 2023) and Willow (CRYPTO 2025). Our core technical contribution is a novel application of recent advances in batched threshold encryption, which enables succinct server–committee interaction while preserving security and correctness. Beyond asymptotic improvements over prior works, Lighthouse yields substantial concrete efficiency gains: For an aggregation with 1024 clients, we reduce server-to-committee communication by over 100× and committee-to-server communication by over 300× compared to Flamingo and Willow. Also, we present an extension that supports dynamic client participation, a critical requirement for practical deployments at scale, while preserving the asymptotic and concrete efficiency of the static protocol for clients.

Assumption-Free Fuzzy PSI via Predicate Encryption

Erik-Oliver Blass, Airbus; Guevara Noubir, Northeastern University

Available Media

We present the first protocol for efficient Fuzzy Private Set Intersection (PSI) that achieves linear communication complexity, does not depend on restrictive assumptions on the distribution of party inputs, and abstains from inefficient fully homomorphic encryption. Specifically, our protocol enables two parties to compute all pairs of elements from their respective sets that are within a given Hamming distance, without constraints on how these sets are structured. Our key insight is that securely computing the (threshold) Hamming distance between two inputs can be reduced to securely computing their inner product. Leveraging this reduction, we construct a Fuzzy PSI protocol using recent techniques for inner-product predicate encryption. To enable the use of predicate encryption in our setting, we establish that these predicate encryption schemes only require a weak notion of simulation security. We also demonstrate how their internal key derivation can be efficiently distributed without a trusted third party.

As a result, our Fuzzy PSI on top of predicate encryption achieves optimal linear communication complexity for arbitrary input distributions. Our implementation validates its feasibility and demonstrates improved performance over the most closely related work.

Track 4

Privacy Anonymity 3

Session Chair: Yanjiao Chen, Zhejiang University

Grand Ballroom I-IV

LPG: Raise Your Location Privacy Game in Direct-to-Cell LEO Satellite Networks

Quan Shi, National University of Singapore; Liying Wang, Peking University; Prosanta Gope, University of Sheffield; Qi Liang and Haowen Wang, Beijing University of Posts and Telecommunications; Qirui Liu, Peking University; Chenren Xu, Peking University and Key Laboratory of High Confidence Software Technologies, Ministry of Education (PKU); Shangguang Wang and Qing Li, Beijing University of Posts and Telecommunications; Biplab Sikdar, National University of Singapore

Distinguished Paper Award Winner

Available Media

Multi-tenant direct-to-cell (D2C) Low Earth Orbit (LEO) satellite networks pose significant risks to users' location privacy by linking Mobile Network Operator (MNO)- managed identities with Satellite Network Operator (SNO)- visible locations. Existing privacy solutions are ill-suited to the resource-constrained hardware and orbital dynamics of these satellite environments. We present LPG (Location Privacy Game), the first protocol-layer solution offering user-configurable location privacy for D2C LEO. LPG achieves this via identity-location decoupling: SNOs provide connectivity without visibility of user identity, while MNOs manage service and billing without access to precise location information. LPG enables offline secure authentication and key agreement without revealing user identity to satellites, supports user-configurable location disclosure at chosen geographic granularity for essential service needs, and ensures fair billing between MNOs and SNOs through privacy-preserving settlement. Our implementation on a real-world in-orbit LEO satellite and commercial mobile phones demonstrates that LPG is practical and viable in resource-constrained, highly-dynamic LEO environments.

Security and Privacy Analysis of Tile's Location Tracking Protocol

Akshaya Kumar, Anna Raymaker, and Michael A. Specter, Georgia Institute of Technology

Available Media

We conduct the first comprehensive security analysis of Tile, the second most popular crowd-sourced location-tracking service behind Apple's AirTags. We identify several exploitable vulnerabilities and design flaws, disproving many of the platform's claimed security and privacy guarantees: Tile's servers can persistently learn the location of all users and tags, unprivileged adversaries can track users through Bluetooth advertisements emitted by Tile's devices, and Tile's anti-theft mode is easily subverted.

Despite its wide deployment—millions of users, devices, and purpose-built hardware tags—Tile provides no formal description of its protocol or threat model. Worse, Tile intentionally weakens its antistalking features to support an antitheft use-case, and relies on a novel "accountability" mechanism to punish those abusing the system to stalk victims.

We examine Tile's accountability mechanism, a unique feature of independent interest; no other provider attempts to guarantee accountability. While an ideal accountability mechanism may disincentivize abuse in crowd-sourced location tracking protocols, we show that Tile's implementation is subvertible and introduces new exploitable vulnerabilities. We conclude with a discussion on the need for new, formal definitions of accountability in this setting.

Revelio: A Global Study of VoIP Blocking

Arjun Gupta, IIT Bombay; Roya Ensafi, University of Michigan; Piyush Kumar, IIT Delhi; Devashish Gosain, IIT Bombay

Available Media

Online censorship has substantially increased over the past decade. While prior work on censorship detection has largely focused on the web, the landscape and mechanisms of VoIP censorship remain poorly understood.

This work presents the first global study of VoIP blocking by introducing Revelio, a system that systematically detects and comprehensively analyzes VoIP censorship at scale. Unlike the only prior work, which is limited to two countries and requires authors to travel to conduct measurements, Revelio remotely uncovers VoIP blocking by carefully crafting probe packets and strategically leveraging IP address space within target countries, enabling detection of otherwise invisible VoIP censorship.

Using Revelio, we studied 71 countries, covering 11 widely used VoIP applications. Our findings reveal evidence of VoIP blocking in 19 countries, primarily through targeted interference with the STUN traffic. By studying packet traces, we reverse-engineer censorship behavior and uncover four classes and six subtypes of middleboxes that trigger on unique application-specific STUN packets. We further analyze domain-level blocking at the HTTP and HTTPS layers, revealing cases where only VoIP calling is blocked while messaging remains accessible.

Overall, our study advances the understanding of the privacy and censorship community by providing the first comprehensive, global view of an understudied yet highly prevalent form of censorship. We release our code to support reproducibility and future research.

Towards Practical Few-shot Multi-tab Website Fingerprinting

Lin Liu and Ziling Wei, National University of Defense Technology; Zhuotao Liu and Xinhao Deng, Tsinghua University; Zixuan Dong and Shuhui Chen, National University of Defense Technology

Available Media

Website fingerprinting (WF) attacks can infer the visited websites to deanonymize Tor networks by analyzing encrypted traffic patterns. Recent few-shot WF methods reduce reliance on large-scale data collection, yet they predominantly formulate WF as a single-label classification task and rely on meta-learning episodes that assume disjoint label sets and stable embedding spaces. These assumptions break down in the realistic multi-tab browsing, where traffic from multiple websites interleaves within a single observation window and the number of concurrent tabs is unknown. Meanwhile, the label space grows exponentially as the monitored set expands, making existing multi-tab WF methods costly to update. To address these challenges, we propose MMF, a novel framework for few-shot multi-tab WF. MMF shifts the meta-learning objective from single-label classification to support-guided presence detection. In each episode, we pair a mixed multi-tab query trace with a small set of single-tab support traces for each monitored website, and generate class-specific features by feature reweighting to decide which websites are present. This detection-centric formulation yields well-defined few-shot tasks in multi-tab scenarios and enables MMF to detect previously unseen websites from limited traces. We evaluate MMF on established public datasets and a new real-world dataset collected under varied browsing conditions. Results demonstrate that MMF consistently outperforms state-of-the-art multi-tab WF attacks across all settings. Notably, in the 5-shot scenario, MMF achieves improvements of up to 300% in Novel Precision@k, highlighting its strong capability for few-shot detection in dynamically growing website sets.

PrivacyShield: Relaying BLE Beacons to Counter Unsolicited Tracking

Florian Hofhammer, EPFL; Daniele Antonioli, EURECOM; Mathias Payer, EPFL

Available Media

Offline finding networks such as Apple's Find My, Google's Find My Device, or Samsung's SmartThings Find are frequently abused to stalk unsuspecting victims. These networks allow users to attach small, cheap tags to items to locate them if they are lost. The tags announce their presence via Bluetooth Low Energy (BLE) beacons, and nearby Internet-connected devices such as smartphones report their location to the finding network. However, the low price and easy-to-hide footprint of offline finding tags makes them appealing to malicious actors, who place tags on their unwitting victims. Nearby devices or even the victim's own device then unknowingly report the victim's location to the stalker.

We analyze the anti-stalking measures put in place by offline finding networks with a focus on Apple's Find My and Google's Find My Device. We show how malicious actors can bypass those measures and propose PrivacyShield, a novel relay network protecting stalking victims. Our network takes advantage of the fact that offline finding BLE beacons are unauthenticated and can be relayed to arbitrary locations. Relayed beacons cause third-party devices to report incorrect locations to the finding network, obfuscating the victim's location. We demonstrate PrivacyShield's effectiveness in masking a tag's location, and show the robustness of the system against attempts to thwart its usage. Then, we suggest practical recommendations for offline finding network providers to improve stalking protection.

Technical Analysis of the Geedge Networks Firewall Source Code Leak

Anna Ablove, University of Michigan; Johnnie Walker, GFW Report; Ben Wolin, University of Michigan; Niklas Niere and Felix Lange, Paderborn University; Aaron Ortwein, Armin Huremagic, and Richa Priyanka, University of Michigan; Ali Zohaib and Jade Sheffey, University of Massachusetts Amherst; Nico Heitmann, Paderborn University; J. Alex Halderman, University of Michigan; Juraj Somorovsky, Paderborn University; Amir Houmansadr, University of Massachusetts Amherst; Roya Ensafi, University of Michigan; Mingshi Wu, GFW Report; Eric Wustrow, University of Colorado Boulder

Available Media

In September 2025, over 100K internal documents (including code, communications, etc.) from Geedge Networks, a Chinese DPI company with ties to the Great Firewall of China, were leaked to the public. In this paper, we analyze the source code from this leak, focusing on Geedge Networks' flagship product, the Tiangou Secure Gateway (TSG) firewall. Working across multiple repositories, we successfully build and run a local copy of TSG---revealing key aspects of its architecture, including the protocols it is capable of parsing and the format of blocking rules used to censor sites, proxies, and other resources. Finally, we extract several fingerprints from TSG, including custom random number generators and parsing idiosyncrasies that allow us to identify its use and similar deployments in the Great Firewall of China.

This is the first time that the source code of a commercial DPI has been leaked, and our work is the first code analysis of a core firewall component used in national censorship infrastructure. This unprecedented investigation offers insights that can assist circumvention developers and Internet security researchers in further understanding the capabilities and limitations of modern censorship technology.

Provable Secure Steganography Based on Adaptive Dynamic Sampling

Kaiyi Pang and Minhao Bai, Tsinghua University

Available Media

The security of private communication is increasingly at risk due to widespread surveillance. Steganography, a technique for embedding secret messages within innocuous carriers, enables covert communication over monitored channels. Provably Secure Steganography (PSS), which ensures computational indistinguishability between the normal model output and steganography output, is the state-of-the-art in this field. However, current PSS methods often require obtaining the explicit distributions of the model. In this paper, we propose a provably secure steganography scheme that only requires a model API that accepts a seed as input. Our core mechanism involves sampling a candidate set of tokens and constructing a map from possible message bit strings to these tokens. The output token is selected by applying this mapping to the real secret message, which provably preserves the original model's distribution. To ensure correct decoding, we address collision cases, where multiple candidate messages map to the same token, by maintaining and strategically expanding a dynamic collision set within a bounded size range. Extensive evaluations of three real-world datasets and three large language models demonstrate that our sampling-based method is comparable with existing PSS methods in efficiency and capacity.

From Symmetry toward Weak Asymmetry: Secure Steganography under the Receiver's Partial Channel Knowledge

Guorui Liao, Shuyu Gan, Jinshuai Yang, Yongfeng Huang, and Zhongyi Huang, Tsinghua University

Available Media

Provably secure steganography enables covert communication by disguising secret messages as normal communication content while guaranteeing computational indistinguishability. Recent work typically assumes that the sender can construct stegotext based on explicit conditional distributions, while the receiver either has access to the same distributions for decoding or has no such access at all. In many realistic communication settings, however, the receiver's access lies between these two extremes, a regime that is not captured by either assumption. We therefore introduce and study a new setting, called weakly asymmetric steganography, in which the sender has access to the full conditional distributions used for stegotext construction, while the receiver only has partial knowledge derived from them. We design provably secure steganography schemes under several representative forms of receiver-side knowledge. Specifically, we consider binary alphabets in which the receiver only knows whether a symbol's probability exceeds a given threshold, as well as general alphabets where the receiver only observes a possibly partial ranking of symbols by probability. We evaluate our rank-based constructions in language-model-based text steganography under both limited receiver-side knowledge and precision-induced distribution mismatch. The results show that weakly asymmetric steganography can achieve non-trivial embedding capacity, and further suggest the potential of this setting in distribution-mismatch scenarios.

Track 5

Enigma Track: Hardening AI

Session Chair: Will Enck, North Carolina State University

Dover Room

When Truth Is Undefined: Why LLM Evaluation Breaks Down in Practice

Thursday, 11:00 am–11:30 am

Vaishali Vinay, Microsoft

Most LLM evaluation pipelines quietly assume something we rarely have in production: a stable ground truth. Benchmarks treat correctness as singular. Human annotation treats disagreement as noise. Metrics assume yesterday's distribution still represents today's reality. In practice, these assumptions often fail. Real-world deployments operate in environments where multiple answers may be valid, labels evolve faster than evaluation datasets, and models that improve on benchmarks can still regress in the decisions they influence.

This talk draws from production LLM evaluation workflows to examine recurring failure modes: annotator disagreement misinterpreted as model error, metric drift mistaken for model regression, benchmark overfitting presented as progress, and evaluations that lose meaning as tasks, users, and environments evolve.

Rather than proposing another benchmark, we argue that evaluation should be treated as decision-making under uncertainty. We present practical patterns that help organizations assess model quality when ground truth is incomplete, contested, or constantly changing.

Vaishali Vinay is a Senior Data & Applied Scientist at Microsoft, where she works on the design, deployment, and evaluation of large-scale AI systems for security and decision-support. With over a decade of experience in applied AI and cybersecurity, she has led and contributed to production systems for threat detection, alert correlation, and AI-driven efficacy analysis. Her work involves human-in-the-loop evaluation, longitudinal quality measurement, and assessing model behavior in real-world operational settings. She has authored multiple patents and publications and has delivered invited talks on applied AI and security. Her work has directly informed enterprise-scale security operations and product decisions.

Breaking the Lethal Trifecta: Secure Agentic Computing via Causal Isolation Architecture

Thursday, 11:30 am–12:00 pm

David Tsao, Notion Labs, Inc.

Available Media

Most LLM agents over private data have a dangerous shape: the same model reads untrusted instructions, sees sensitive context, and speaks to the outside world. It is like handing a press secretary every private email, incident note, and salary spreadsheet, then hoping a policy memo keeps secrets safe.

Causal Isolation Architecture treats leakage as a systems design failure, not a prompt-writing failure. It splits the lethal trifecta across isolated components, passes only typed abstractions across boundaries, and uses a rule-of-two release judge to preserve useful private-derived guidance while stripping exact facts based on requester trust and source sensitivity.

The key lesson: safety without utility becomes a refusal machine. In our bounded real-model matrix, claim-eligible OpenAI and Anthropic rows showed 0.0% observed secure-path leakage with 83-100% measured utility; matching single-agent baselines leaked private facts. The talk shows the demo, tradeoffs, and limits.

David Tsao is Notion's Chief Information Security Officer (CISO), leading the teams responsible for Platform Security, IT, and Security Engineering. With 20+ years of experience building security culture and risk-based security programs across high-growth SaaS and large enterprises, he partners closely with company leadership to strengthen Notion's security posture and protect customers' data. Prior to Notion, David held CISO and senior security leadership roles at Instacart, Marqeta, BYTON, and Veeva Systems, and spent more than a decade at Gilead Sciences building enterprise security and privacy programs; he is also a CISSP and advises startups as a venture advisor.

12:30 pm–2:00 pm

Symposium Luncheon

Harborside Ballroom

2:00 pm–3:30 pm

Track 1

Privacy Anonymity 4

Session Chair: Maryna Veksler, Virginia Commonwealth University

Grand Ballroom VII-X

Do You Need a Receipt? Anonymous Credential Revocation at Continental Scale via Private Record Certification

Kasra EdalatNejad, Sebastian Faust, Jonas Hofmann, Philipp-Florens Lehwalder, and Thomas Schneider, Technical University of Darmstadt

Distinguished Artifact Award Winner

Available Media

A key challenge in digital credential systems is revocation, that is, the ability to invalidate credentials post-issuance and verify their status upon presentation. While anonymous credentials enhance privacy over classical credentials (e.g., by providing unlinkability), they complicate revocation. Existing revocation schemes for anonymous credentials often suffer from high client or verifier computation, long delays before revocation takes effect (e.g., epoch-based settings), or require updates to all users with each revocation.

We present an efficient, real-time revocation system for anonymous credentials with decentralized revocation authorities based on a novel primitive called Private Record Certification (PRC). PRC enables users to obtain a certificate for a record stored in a server-managed database without the servers learning which record was requested. This primitive is of independent interest, and we construct it by combining techniques from private information retrieval and secure multi-party computation. Our revocation scheme outsources its computation to the revocation authorities and has minimal overhead for clients and verifiers, while ensuring the communication costs are sublinear in the number of credentials. We build a prototype and demonstrate that our system achieves sub-second real-time latency for PRC requests, at a scale of over 1 billion credentials, with an online operational cost of 2.5$ per server for processing 1 million PRC requests.

Remise: Authorized Anonymous Communication Systems

Rohan Ravi, Paritosh Shukla, and Adithya Vadapalli, IIT Kanpur

Available Media

We present Remise, a two-server authorized anonymous communication system built on Distributed Oblivious RAM (DORAM). Remise supports two modes of operation: (i) an anonymous bulletin board, where messages are publicly revealed at the end of each epoch without linking senders to messages, and (ii) anonymous communication channels, where messages remain secret-shared, and writers selectively grant and revoke read access to chosen readers. In both modes, the two servers execute read and write operations without learning which indices are accessed, which authorization tokens are used, or which relationships exist between writers and readers, assuming no collusion between the servers. A central contribution of Remise is a lightweight and efficient access-control mechanism. Authorization tokens are maintained in secret-shared form across the two servers, enabling oblivious verification and updates, while preventing leakage even under client–server collusion. Unlike prior DPF-based systems, Remise provides built-in auditing by having the semi-honest servers generate standard-basis vector shares internally, eliminating the need for server-side DPF validity checks. We implement a prototype of Remise and evaluate it under realistic network conditions. Our experiments show an 80× improvement in online server time for request validity check when compared to PACL (IEEE S&amp;P 2023) for large database size.

SophOMR: Improved Oblivious Message Retrieval from SIMD-Aware Homomorphic Compression

Keewoo Lee, Ethereum Foundation; Yongdong Yeo, Seoul National University

Available Media

Privacy-preserving blockchains and private messaging services that ensure receiver-privacy face a significant UX challenge: each client must scan every payload posted on the public bulletin board to avoid missing messages intended for them. Oblivious Message Retrieval (OMR) addresses this issue by securely outsourcing this expensive scanning process to a service provider using Homomorphic Encryption (HE).

In this work, we propose a new OMR scheme that substantially improves upon the previous state-of-the-art, PerfOMR (USENIX Security'24). Our implementation demonstrates reductions of 3.4x in runtime, 2.2x in digest size, and 1.5x in key size, in a scenario with 65536 payloads (each 612 bytes), of which up to 50 are pertinent. At the core of these improvements is a new homomorphic compression mechanism, where ciphertexts of length proportional to the number of total payloads are compressed into a digest whose length is proportional to the upper bound on the number of pertinent payloads. Unlike previous approaches, our scheme fully exploits the native homomorphic SIMD structure of the underlying HE scheme, significantly enhancing efficiency. In the setting described above, our compression scheme achieves 7.5x speedup compared to PerfOMR.

Sliced Rényi Pufferfish Privacy: Tractable Privatization Mechanism and Private Learning with Gradient Clipping

Tao Zhang and Yevgeniy Vorobeychik, Washington University in St. Louis

Available Media

We study the design of a privatization mechanism and privacy accounting in the Pufferfish Privacy (PP) family. Specifically, motivated by the curse of dimensionality and lack of practical composition tools for iterative learning in the recent Rényi Pufferfish Privacy (RPP) framework, we propose Sliced Rényi Pufferfish Privacy (SRPP). SRPP preserves PP/RPP semantics (customizable secrets with probability-aware secret–dataset relationships) while replacing high-dimensional Rényi divergence with projection-based quantification via two sliced measures, Average SRPP and Joint SRPP. We develop sliced Wasserstein mechanisms, yielding sound SRPP certificates and closed-form Gaussian noise calibration. For iterative learning systems, we introduce an SRPP-SGD scheme with gradient clipping and new accountants based on History-Uniform Caps (HUC) and a subsampling-aware variant (sa-HUC), enabling decompose-then-compose privatization and additive composition under a common slicing geometry. Experiments on static and iterative privatization show that the proposed framework exhibits favorable privacy–utility trade-offs, as well as practical scalability.

Nudge: A Private Recommendations Engine

Alexandra Henzinger, MIT; Emma Dauterman, MIT and Stanford; Henry Corrigan-Gibbs, MIT; Dan Boneh, Stanford

Available Media

Nudge is a recommender system with cryptographic privacy. A Nudge deployment consists of three infrastructure servers and many users, who retrieve/rate items from a large data set (e.g., videos, posts, businesses). Periodically, the Nudge servers collect ratings from users in secret-shared form, then run a three-party computation to train a lightweight recommender model on users' private ratings. Finally, the servers deliver personalized recommendations to each user. At every step, Nudge reveals nothing to the servers about any user's preferences beyond the aggregate model itself. User privacy holds against an adversary that compromises the entire secret state of one server. The technical core of Nudge is a new, three-party protocol for matrix factorization. On the Netflix data set with half a million users and ten thousand items, Nudge (running on three 192-core servers on a local-area network) privately learns a recommender model in 50 mins with 40 GB of server-to-server communication. On a standard quality benchmark (nDCG@20), Nudge scores 0.29 out of 1.0, on par with non-private matrix factorization and just shy of non-private neural recommenders, which score 0.31.

B-Privacy: Defining and Enforcing Privacy in Weighted Voting

Samuel Breckenridge, Dani Vilardell, and Andrés Fábrega, Cornell Tech, IC3; Amy Zhao, Ava Labs, IC3; Patrick McCorry, Arbitrum Foundation; Ari Juels, Cornell Tech, IC3

Available Media

In traditional, one-vote-per-person voting systems, privacy equates with ballot secrecy: voting tallies are published, but individual voters' choices are concealed.

Voting systems that weight votes in proportion to token holdings, though, are now prevalent in cryptocurrency and web3 systems. We show that these weighted-voting systems overturn existing notions of voter privacy. Our experiments demonstrate that even with secret ballots, publishing raw tallies often reveals voters' choices.

Weighted voting thus requires a new framework for privacy. We introduce a notion called B-privacy whose basis is bribery, a key problem in voting systems today. B-privacy captures the economic cost to an adversary of bribing voters based on revealed voting tallies.

We propose a mechanism to boost B-privacy by noising voting tallies. We prove bounds on its tradeoff between B-privacy and transparency, meaning reported-tally accuracy. We show experimentally across 2,503 proposals in 27 Decentralized Autonomous Organization (DAOs) that, with minimal transparency degradation, our mechanism raises B-privacy by a geometric mean factor of 3.5×.

Our work offers the first principled, practical, systemic guidance for weighted-voting systems, complementing existing approaches that focus on ballot secrecy.

Unbalanced Fuzzy Private Set Intersection for L_infinity Distance: Achieving Sublinear Communication with Large Set Size

Shengzhe Meng and Xiaodong Wang, Tsinghua University; Xv Zhou, Beihang University; Bei Liang, Beijing Institute of Mathematical Sciences and Applications

Available Media

Fuzzy private set intersection (PSI) is a cryptographic protocol that enables two parties to compute the intersection of their sets under approximate matching, with variants including standard fuzzy PSI and fuzzy PSI with sender privacy (PSI-SP). Although recent advances have led to efficient fuzzy PSI protocols, most are designed for the balanced case where both sets are of similar size. In practice, however, many applications involve highly unbalanced sets (e.g., where the receiver's set is much larger than the sender's, or vice versa). This work focuses on unbalanced fuzzy PSI for the l∞ metric. We observe that communication in existing protocols is dominated by the transmission of oblivious key-value stores (OKVS), especially when set sizes are imbalanced. This overhead can be reduced using batch private information retrieval (BatchPIR) if the OKVS is sparse. However, such optimization requires spatial hashing with specific properties, and few existing spatial hashing schemes satisfy these requirements.

In this work, we reformulate spatial hashing and propose two new schemes: one non-interactive and one interactive, each suited to different input set conditions. Based on these, we design two unbalanced fuzzy PSI protocols that combine sparse OKVS with BatchPIR to achieve sublinear communication in the size of the larger set. The first protocol is suitable for scenarios where the receiver holds a larger set, while the second is designed for cases where the sender possesses more items. Our protocols significantly outperform state-of-the-art in communication and runtime. For example, in a 100 Mbps network with parameters (N, M, d, σ) = (220, 25, 2, 10), our protocol based on our non-interactive spatial hashing reduces communication from 16,128 MB (Baarsen and Pu, Eurocrypto'24) to 0.35 MB. Furthermore, our fuzzy PSI protocol, which utilizes our interactive spatial hashing approach, achieves at least 31× faster online runtime and 1762× lower communication than Gao et al. (Asiacrypt'25). For our fuzzy PSI protocol with sender privacy, we outperform Piske et al. (CCS'25) with at least 4× faster online runtime and 6× lower communication.

sigma-rs: A Modular Approach for Keyed-Verification Anonymous Credentials

Michele Orrù, CNRS; Lindsey Tulloch, Tor Project, University of Waterloo; Victor Snyder-Graf, RISC Zero; Ian Goldberg, University of Waterloo

Available Media

We introduce a new software stack in Rust aimed at simplifying constructions and deployments of protocols based on modern anonymous credential systems.

The stack, called sigma-rs, through its layered design, abstracts cryptographic complexity while remaining flexible enough to support a range of credential schemes, proofs, and access policies. It emphasizes misuse resistance via type safety, domain separation, and prover-state discipline, and supports side-channel-aware constant-time strategies.

We evaluate practicality through re-implementations of Tor's Lox bridge distribution protocols and of user authentication in the Open Observatory for Network Interference.

Track 2

Hardware Security 3

Session Chair: Qinhong Jiang, The Hong Kong Polytechnic University

Grand Ballroom VI

Design and Implementation of a Physical Implant Attack on the Boeing 737

Sam Crow, UC San Diego; Stephen Checkoway, Oberlin College; Patrick Mercier, Pat Pannuto, Stefan Savage, and Aaron Schulman, UC San Diego

Available Media

We explore a new kind of threat model for aviation cybersecurity—one in which an adversary has temporary physical access to an airframe. We argue why such attacks are practically feasible and explain how the design of existing avionics systems, their interconnects, and their maintenance architecture make such threats of particular concern. Using the Boeing 737 as an example, we develop and demonstrate a small, programmable hardware implant that can be quickly inserted into an existing maintenance socket with roughly 60 seconds of access on the ground. By carefully manipulating physical ARINC 429 bus signals, this device can create an undetected "attacker-in-the-middle" (AITM) capability between the aircraft's flight management computer (FMC) and multipurpose control display unit (MCDU). We explore the options for addressing this class of attack and, in particular, show why transformer-coupled buses such as MIL-STD-1553, are far more challenging to manipulate in this manner.

FIRA: Enabling Automatic Forensic Investigation of Unmanned Aerial Vehicles

Yizhi Huang, Georgia Institute of Technology; David Oygenblik, Georgia Tech; Runze Zhang, Mingxuan Yao, Muhammad Ibrahim, Burak Sahin, Haichuan Xu, Saman Zonouz, and Brendan Saltaformaggio, Georgia Institute of Technology

Available Media

In dynamic environments, unmanned aerial vehicles (UAVs) often utilize online learning to refine their machine learning (ML) model's decision boundaries for improved performance. Unfortunately, when the UAV becomes irrecoverable or unavailable (e.g., a crash), a forensic investigator would be left helpless to determine if the UAV's online learning caused the crash. This paper proposes a novel forensic technique, called FIRA, that can establish causal connections from ML models to UAV system components. FIRA sends back in-flight online learning updates and telemetry data (even when bandwidth is limited) and determines whether the crash can be attributed to the online learning model. We applied FIRA to 48 UAV crash scenarios using two widely adopted UAV control programs: PX4 and ArduPilot. Across four types of UAV missions, FIRA investigated 12 accidents (each) in which a backdoored online learning model was the cause of the crash, and FIRA was able to correctly attribute the model to the crash with 95.8% accuracy.

BADControl: Backdoor Attacks Against Control Systems

Luis Burbano, University of California, Santa Cruz; Hampei Sasahara, The University of Tokyo; Ruoyu Song and Z. Berkay Celik, Purdue University; Alvaro A. Cardenas, University of California, Santa Cruz

Available Media

We introduce BADCONTROL, the first backdoor attack against low-level controllers that uses physical triggers. The attack poisons operational data to implant a vulnerability that can be activated by an exogenous signal from the environment, such as a specific driving maneuver or adversarial road patches within autonomous driving applications. BADCONTROL solves a constrained optimization problem by using a projected gradient ascent to modify the data, maximizing the frequency response of the controlled system at a target frequency. This method differs from backdoor attacks against Deep Learning (DL) and Reinforcement Learning (RL) models, which manipulate high-dimensional model inputs or reward functions. We additionally propose two defenses: one based on regularization and one based on robust optimization, to limit the worst-case amplification of trigger signals. This is achieved by converting infinite poisoning scenarios into a single, tractable optimization problem via a specialized mathematical transformation. We evaluate BADCONTROL on Proportional-Integral-Derivative (PID) and Linear-Quadratic-Regulator (LQR) controllers through simulations and physical experiments. In the adaptive cruise control scenario, we achieve a 100% crash rate, while in lane-keeping control, the backdoor causes the victim vehicle to steer 62% into the opposing lane, compared to 0% in both cases without a backdoor. By contrast, a state-of-the-art falsification framework for autonomous vehicles identifies only a single crash instance over 30 trials, underscoring its stealthiness.

A Cuckoo in the Nest: Multi‑Stage, Multi‑Identifier Hijacking in BACnet/SC

Qiguang Zhang, Junzhou Luo, and Zhen Ling, Southeast University; Yue Zhang, Shandong University and Quancheng Laboratory; Kaizheng Liu and Chongqing Lei, Southeast University; Matthew Harper and Xinwen Fu, University of Massachusetts Lowell

Available Media

To address the well-known security limitations of legacy Building Automation and Control Network (BACnet), BACnet Secure Connect (BACnet/SC) introduces mutually authenticated WebSocket Secure (WSS) channels and mandates PKI-based certificate management. Despite these protections, we identify a fundamental identifier-binding failure in BACnet/SC's security model: X.509 certificate authentication is not cryptographically bound to the logical identifiers used for connection management (UUID) and message forwarding (VMAC). Through a systematic analysis of BACnet/SC connection state machines, we show that this decoupling enables a multi-stage, multi-identifier hijacking attack, in which malicious roles progressively displace a legitimate device's authenticated connection state and persistently intercept its traffic. We term this attack the Cuckoo Attack and validate it across the official BACnet/SC Reference Stack, the open-source BACnet-Stack, and commercial building automation platforms from major vendors, including Siemens, Johnson Controls, Honeywell, and Carrier. We further propose mitigation measures to address this vulnerability. Our findings provide the first systematic characterization of an authentication-identifier decoupling flaw in BACnet/SC, highlight a broader class of identifier-binding risks in stateful secure protocols, and have been presented to and acknowledged by ASHRAE SSPC 135, the standards committee responsible for BACnet.

ARTO: Efficient Execution Integrity Attestation for Real-Time Operation of Cyber-Physical Systems

Ruizhe Zhao, Cong Sun, and Zongzhen Li, State Key Laboratory of Integrated Services Networks, School of Cyber Engineering, Xidian University; Tiantian Wang, Northeastern University; Yunbo Wang, State Key Laboratory of Integrated Services Networks, School of Cyber Engineering, Xidian University

Available Media

Real-time embedded systems are prevalent in cyber-physical applications such as drones and autonomous vehicles. Due to their high complexity, the environmental uncertainties, and the openness to remote control, these systems are vulnerable to control-flow hijacking and data-only attacks that compromise runtime integrity and reliability. Traditional control-flow integrity (CFI) and data-flow integrity (DFI) approaches impose excessive overhead, while the control-flow attestation (CFA) allows detection latency incompatible with real-time constraints. This paper presents ARTO, the first control-flow and data-flow protection approach that integrates the control-flow attestation to provide strong security with competitive runtime overhead. Through the prover-side path segmentation and hash-based validation result caching, ARTO provides partial context-sensitive protection at minimal runtime cost, offloading complete context-sensitive detection to the remote verifier. Furthermore, ARTO implements an equivalence-class-based data-flow protection that combines the address-based target check for memory reads with the value-based checks to protect critical variables, stronger than the SOTA operation integrity enforcement in validating the read source for critical variables. Extensive evaluations on robotic vehicle autopilots confirm ARTO's efficacy in mitigating both control-flow and data-only attacks while maintaining runtime overhead within real-time limits, outperforming SOTA CFA and operation-integrity enforcement methods in operation execution time.

TAT: Attesting Trajectory Integrity of Industrial Robotic Arms

Chengtao Yao, Chengcheng Zhao, Peng Cheng, and Jiming Chen, Zhejiang University

Available Media

Industrial robotic arms are central to modern manufacturing, with broad deployment in critical domains. Motion is a primary security concern, as it is a fundamental capability of robotic arms, and adversarial manipulation (e.g., altering production logic, positioning, or dynamics) can lead to product defects or physical damage. Remote attestation is a promising mechanism for verifying execution integrity. However, existing approaches focus on control-flow or data-flow properties and fail to capture motion semantics, limiting their ability to adequately verify the physical execution of robotic arms.

This paper presents Trajectory Integrity (TI) as a new security property that ensures a robotic arm's motion conforms to its intended path. To enforce TI, we design TAT, a minimally invasive attestation framework that leverages a Timed Motion Event Graph to capture motion semantics and combines event and joint measurements to verify actual motion. We implement a hardware-software prototype of TAT on an open-source robotic arm platform. Evaluation on real-world task programs shows that TAT incurs at most 2.30% memory overhead and 0.14% execution time overhead, demonstrating its performance and practicality. Furthermore, its attestation capability is evaluated under diverse motion-related parameter modifications, confirming its effectiveness in trajectory integrity attestation.

BLE Theft Auto: Evaluating the Security of Aftermarket BLE-based Automotive Remote Control Systems

Jerry Yu, Yibo Wei, Sumanth Rao, Mohak Vaswani, and Jefferson Chien, UC San Diego; Christian Dameff, UC San Diego Health; Nishant Bhaskar and Aaron Schulman, UC San Diego

Available Media

Vehicles are increasingly remotely controllable from smartphone apps that function like keyfobs: unlocking a vehicle's doors and even starting its engine. These systems use Bluetooth Low Energy (BLE) for offline remote control. However, they do not use BLE's built-in security; instead they use proprietary application-layer protocols, introducing a new attack surface. This is especially concerning for aftermarket BLE remote control systems, because they have not received scrutiny by security researchers. In this paper, we evaluate feasible BLE-based attacks on aftermarket vehicle remote control systems. We discovered six aftermarket BLE remote control systems in use today, and analyzed their application-layer BLE security. We found critical vulnerabilities in three systems, exposing what we estimate to be 1.4 million aftermarket BLE control systems to attack. We show that targeted BLE remote attacks are practical: an attacker can use crowdsourced Bluetooth scanning databases to target specific vehicles with vulnerable systems installed. Our disclosures lead all vendors to patch improved app-layer BLE security into their systems, affecting many vehicle owners that did not even know they had a BLE remote control system installed.

SoK: Security of Cyber-physical Systems Under Intentional Electromagnetic Interference Attacks

Qinhong Jiang, The Hong Kong Polytechnic University; Yan Long, The Hong Kong University of Science and Technology (Guangzhou); Youqian Zhang, The Hong Kong Polytechnic University; Chen Yan and Xiaoyu Ji, Zhejiang University; Xiapu Luo, The Hong Kong Polytechnic University; Kevin Fu, Northeastern University; Jiannong Cao, The Hong Kong Polytechnical University; Wenyuan Xu, Zhejiang University

Available Media

Falsifying electrical signals in computer systems—the gateway between the physical and digital worlds—intentional electromagnetic interference (IEMI) attacks have become increasingly pervasive and damaging to cyber-physical systems due to their ability to disrupt or control a wide range of safety- and security-critical applications. Existing studies of IEMI attacks are often highly device-specific and exploit disparate, insufficiently compared attack vectors. The absence of a unified, model-based understanding of IEMI vulnerabilities hinders both transferable security assessments and effective cross-disciplinary collaboration toward deployable protections. To address this gap, this work analyzes over 80 instances of IEMI attacks and defenses to provide an analytical framework that models how adversaries achieve IEMI coupling and sample manipulation to inject malicious electromagnetic energy that alters hardware behavior and impacts software execution. The primary goal is to move the field beyond exhaustive empirical discovery of vulnerable instances and toward in-depth theoretical analysis and proactive defense strategies applicable to both existing and future cyber-physical systems. In addition to identifying gaps in current IEMI attack and defense research, this work outlines important directions for future work tailored to the needs and roles of different stakeholder communities. To foster future research on IEMI attacks, we are releasing and maintaining an open-source IEMI research database at https://iemi-research-database.github.io/.

Track 3

Network Security 3

Session Chair: Katherine Izhikevich, University of California, San Diego

Grand Ballroom V

Hop: A Modern Transport and Remote Access Protocol

Paul Flammarion, Stanford University; George Hosono, Georgia Institute of Technology; Wilson Nguyen, Laura Bauman, Daniel Rebelsky, and Gerry Wan, Stanford University; David Adrian, Independent; Zakir Durumeric, Stanford University

Available Media

Since SSH's standardization nearly 20 years ago, real-world requirements for a remote access protocol and our understanding of how to build secure cryptographic network protocols have both evolved significantly. In this work, we introduce Hop, a transport and remote access protocol designed to support today's needs. Building on modern cryptographic advances, Hop reduces SSH protocol complexity and overhead while simultaneously addressing many of SSH's shortcomings through a cryptographically-mediated delegation scheme, native host identification based on lessons from TLS and ACME, client authentication for modern enterprise environments, and support for client roaming and intermittent connectivity. We present concrete design requirements for a modern remote access protocol, describe our proposed protocol, and evaluate its performance. We hope that our work encourages discussion of what a modern remote access protocol should look like in the future.

Lost in Encapsulation: Exploiting Open Tunnelling Hosts and Attacking Private Networks

Angelos Beitis and Mathy Vanhoef, DistriNet, KU Leuven

Available Media

Tunnelling protocols form an essential backbone of today's Internet, where they primarily interconnect remote networks. However, many tunnelling hosts do not implement authentication, making them behave as open one-way proxies and, in turn, introducing denial-of-service (DoS) risks.

In this paper, we show that the security risk of open tunnelling hosts extends beyond DoS attacks. First, we study NAT-enabled tunnelling hosts and demonstrate a novel attack that manipulates the NAT's TCP state, enabling an attacker to fingerprint and inject traffic into the co-located private network. We also study tunnelling hosts that translate global IP addresses, and turn them into two-way proxies. We demonstrate how these two-way proxies enable an attacker to: (1) spoof TCP connections; and (2) launch a novel loop DoS attack. Additionally, we show that Layer 2 (L2) tunnelling hosts, specifically GRETAP, can be abused to attack the LAN, and we demonstrate the impact with 5 attacks: leaking the host's global and private MAC addresses, DHCP starvation, DNS poisoning, ARP spoofing, and an infinite loop DoS attack.

To evaluate the prevalence of vulnerable hosts, we conduct 9 Internet-wide scans and identify that 438,280 hosts are potentially vulnerable. Overall, we show that open tunnelling protocols enable new attacks on co-located private networks, and we hope our work encourages the adoption of new defences.

From Mirai to Gorilla: Deep Dive into a Long-Lasting DDoS-for-Hire Botnet

Maarten Weyns, Dario Ferrero, and Stefan Op de Beek, Delft University of Technology (TU Delft); Daniel Wagner, Max Planck Institute for Informatics / DE-CIX; Georgios Smaragdakis and Harm Griffioen, Delft University of Technology (TU Delft)

Available Media

In 2016, the Mirai botnet swept the Internet, ushering in a new era of DDoS attacks. Over the following decade, spinoffs of the Mirai botnet transitioned from simple attack tools into commercial platforms, offering Distributed Denial of Service (DDoS) attacks for Hire. Such platforms enable users to launch large-scale DDoS attacks with minimal technical expertise. One notable example is the Gorilla Botnet, which was operational between Fall 2024 and Summer 2025, an unusually long lifetime compared to similar Mirai-based Botnets.

In this paper, we reverse-engineer the Mirai-based Gorilla Botnet and aim to understand its design, engineering decisions, and marketing strategies to enhance its resilience and success. We investigate its operational characteristics, including the types of attacks it supports, its underlying infrastructure, and the behavior of its bots. We find that Gorilla's longevity stems from targeted improvements, including two software development phases and learning from previous releases, setting it apart from typical Mirai-based botnets. In the process, we analyze the firepower and attack vectors of the Gorilla botnet and characterize the business types of its targets.

TED: Abusing Tunnel Hosts and IPv6 Extension Headers for Pulsing DoS Attacks

Le Gai, Zedong Jia, Lin He, Daguo Cheng, Chentian Wei, and Ying Liu, Tsinghua University

Available Media

IP tunneling mechanisms are widely deployed to facilitate the Internet's transition from IPv4 to IPv6, yet their security implications remain insufficiently scrutinized. In this paper, we present TED, a novel pulsing denial-of-service attack that exploits structural vulnerabilities in IP tunneling and IPv6 Extension Headers (EHs) processing. Unlike prior reflection attacks that depend on application-layer services, TED operates entirely at the network layer, requiring no victim interaction, prolonged traffic accumulation, or protocol-specific dependencies. Delay lines are constructed by TED through nested EHs, capitalizing on the tunnel hosts' blind forwarding logic and lack of deep packet inspection. These artificial delay lines allow attackers to solve the send-time schedule problem, converging asynchronous, low-rate traffic into a destructive, high-magnitude pulse at the victim. Our Internet-wide measurement identified over 1.9 million vulnerable tunnel hosts acting as unwitting relays, including hosts within critical satellite infrastructure. Evaluation results demonstrate that TED achieves normalized pulse amplitude exceeding 180× while successfully evading existing low-rate and pulsing attack detection mechanisms based on specific application layer protocols. We discuss the root causes of these transitional vulnerabilities and possible mitigation strategies for network operators and equipment vendors.

vCause: Efficient and Verifiable Causality Analysis for Cloud-based Endpoint Auditing

Qiyang Song, Qihang Zhou, Xiaoqi Jia, and Zhenyu Song, Institute of Information Engineering, Chinese Academy of Sciences; and School of Cyber Security, University of Chinese Academy of Sciences; Wenbo Jiang, University of Electronic Science and Technology of China; Heqing Huang, Independent Researcher; Yong Liu, Qi An Xin Technology Group Inc.; Dan Meng, Institute of Information Engineering, Chinese Academy of Sciences; and School of Cyber Security, University of Chinese Academy of Sciences

Available Media

In cloud-based endpoint auditing, security administrators often rely on the cloud to perform causality analysis over log-derived versioned provenance graphs to investigate suspicious attack behaviors. However, the cloud may be distrusted or compromised by attackers, potentially manipulating the final causality analysis results. Consequently, administrators may not accurately understand attack behaviors and fail to implement effective countermeasures. This risk underscores the need for a defense scheme to ensure the integrity of causality analysis. While existing tamper-evident logging schemes and trusted execution environments show promise for this task, they are not specifically designed to support causality analysis and thus face inherent security and efficiency limitations.

This paper presents VCAUSE, an efficient and verifiable causality analysis system for cloud-based endpoint auditing. VCAUSE integrates two authenticated data structures: a graph accumulator and a verifiable provenance graph. The data structures enable validation of two critical steps in causality analysis: (i) querying a point-of-interest node on a versioned provenance graph, and (ii) identifying its causally related components. Formal security analysis and experimental evaluation show that VCAUSE can achieve secure and verifiable causality analysis with only <1% computational overhead on endpoints and 3.36% on the cloud.

Toward Understanding the Security Implications in Python Configuration Files

Xinwei Yu, University of Delaware; Zhenkai Zhang, Clemson University; Yuzhe Tang, Syracuse University; Xing Gao, University of Delaware

Available Media

Security incidents targeting open-source software have increased substantially in recent years, yet existing defenses primarily focus on analyzing source code while largely overlooking configuration files that define how software is built, installed, and executed. In this paper, we present an in-depth study of security risks in Python configuration files, analyzing five widely used formats across the project lifecycle. We focus on risks that exploit configuration files to silently redirect dependency resolution to attacker-controlled infrastructure or hijack benign commands. We design an automated framework and identify 39 configuration fields that can be exploited under realistic usage scenarios. We further demonstrate that existing malware detection tools fail to detect most of these risks. As mitigation, we develop ConfigScoper, which not only examines security-relevant fields but also employs multiple vulnerability detectors to identify potential malicious behaviors. We apply ConfigScoper in a measurement study analyzing millions of open-source Python projects on GitHub, and our empirical results show that exploitable risks already exist in real-world projects. Finally, we have responsibly disclosed the identified vulnerabilities to relevant stakeholders.

FRAGJAM: DoS Attacks Using IP Reassembly Congestion

Yepeng Pan and Christian Rossow, CISPA Helmholtz Center for Information Security

Available Media

UDP, one of the major transport protocols for multiple popular services such as DNS and video conferencing, is an important component of today's network. Its reliance on IP fragmentation is known to cause both security and reliability issues. To avoid fragmentation, UDP-based applications hence usually limit the payload size. Currently, Linux's IP fragment reassembling algorithm relies on a per-network-namespace buffer size limit. That is, a classic resource-exhaustion DoS attack against UDP services relying on IP fragmentation is possible. However, the fragility of present real-world services has not yet been thoroughly researched.

In this paper, we examine the practicability of such an IP-fragmentation-based DoS attack against real-world providers of popular UDP-based services, including VPN, video conferencing, and RADIUS. We assess the attack from both the client and service provider perspectives. On the server side, we observe that many real-world service providers fragment the server-to-client traffic with respect to artificially small path MTUs when receiving attacker forged ICMP Fragmentation Needed messages. On the client side, we show the chance of dropping server-to-client fragmented traffic by flooding Linux-based NAT gateways with bogus fragments. Through simulated attacks, we demonstrate that the fragmentation-based DoS attack is realistic, affecting various providers such as Zoom and ExpressVPN. We conduct a comprehensive disclosure to affected parties and suggest possible mitigations.

WAVED: Principled Identification of Off-Path Exploitable Weak Verifications within the TCP/IP Protocol Suite

Yizhou Zhao and Xuewei Feng, Tsinghua University; Min Li, Zhongguancun Laboratory; Ke Xu, Tsinghua University

Available Media

Off-path exploits targeting the fundamental TCP/IP protocol suite pose significant threats to the security of the Internet infrastructure. In particular, weak verifications of received payloads—arising from the lack of reliable information to validate or implementation flaws within the suite—lead to vulnerabilities that attackers can exploit to manipulate traffic, induce data loss, and disrupt services on victim servers. In this paper, we present the first systematic study of these vulnerabilities and introduce WAVED, a framework for identifying off-path exploitable weak verifications within the TCP/IP protocol suite implementation. At the core of WAVED, we develop a flow-, context-, and field-sensitive pointer analysis tailored to the TCP/IP kernel, and construct a Taint Propagation Graph (TPG) to model and trace data flow within the stack. By modeling byte-granularity taint propagation across diverse arithmetic operations, our approach can accurately locate specific input bytes associated with each constraint. Furthermore, direction-sensitive taint information is computed to accurately capture and differentiate the strength of constraints imposed by alternative branch outcomes, thereby significantly outperforming traditional byte-insensitive and direction-insensitive analyses. We evaluate WAVED on IPv4 and IPv6 across Linux 5.15, Linux 6.8, and FreeBSD 14.1. It precisely uncovers weak verifications leading to semantic vulnerabilities in TCP/IP and reveals 14 previously unknown vulnerabilities. We have responsibly disclosed these vulnerabilities to the affected OS vendors and have received acknowledgments from the Linux community.

Track 4

System Security 2

Session Chair: Hong Hu, The Pennsylvania State University

Grand Ballroom I-IV

Download More RAM: Dismantling Windows Operating System Defences with Mischievous Memory

Sam Collins, Tom Chothia, William Burgess, and Marius Muench, The University of Birmingham; David Oswald, Durham University

Distinguished Paper Award Winner

Available Media

Virtualisation-Based Security (VBS) is the cornerstone of modern Windows desktop defences, relied upon by both the operating system and third-party software, with a virtualised secure kernel providing strong security guarantees against even privileged attackers. In this paper, we introduce Download More RAM, software-only memory aliasing attack that breaks these guarantees without physical access. On systems running the most common consumer DIMMs, our attack allows arbitrary memory read/write, letting a privileged user compromise the OS at every level, including the secure kernel, Hypervisor Enforced Code Integrity (HVCI), and all defences it provides. With this access we develop a series of case study attacks targeting VBS-protected processes, Windows Defender, anti-virus & EDR software, and game anti-cheats. Our work breaks the strongest security guarantees offered by the Windows OS, questioning key trust assumptions on such systems. Microsoft have assigned CVE-2026-23670 to our findings and issued a patch that partially mitigates our attack.

Hydrangea: Optimistic Two-Round Partial Synchrony with Improved Fault Resilience

Nibesh Shrestha, Supra Research; Aniket Kate, Supra Research / Purdue University; Kartik Nayak, Duke University

Available Media

Consensus protocols in the partially synchronous setting face a fundamental trade-off: achieving optimal Byzantine fault tolerance requires a good-case latency of at least three rounds, while committing in fewer than three rounds generally entails reduced resilience. Even optimistic protocols such as SBFT (DSN'19), FaB (TDSC'06), and Kudzu achieve an optimistic good-case latency of two rounds under favorable conditions, but only at the cost of reduced fault tolerance.

In this work, we introduce Hydrangea, a partially synchronous Byzantine fault-tolerant state machine replication protocol that combines low latency with improved fault resilience. Let f denote the maximum number of tolerated Byzantine faults, c the maximum number of tolerated crash faults, and k 0 a tunable parameter. For a system of n = 3f + 2c + k + 1 parties, Hydrangea achieves an optimistic good-case latency of two rounds when the total number of faulty parties (Byzantine or crash) is at most p = c + k⁄2 \rfloor. In more adversarial settings, with up to f Byzantine faults and c crash faults, it guarantees a good-case latency of three rounds. We further prove a matching lower bound: no protocol can achieve a two-round optimistic commit under this fault model if p > c + k + 2⁄2 \rfloor.

Our experimental evaluation on geo-distributed deployments demonstrates that Hydrangea consistently achieves substantially lower latency than state-of-the-art protocols in both Byzantine-only and Byzantine–crash fault models, while also delivering modest improvements in throughput.

Secpoline: A Scalable Approach to Build Secure In-Process Syscall Interposers

Ruben Sturm and Anton Schelfhout, DistriNet, KU Leuven; Merve Gülmez, Ericsson Security Research; Adriaan Jacobs and Stijn Volckaert, DistriNet, KU Leuven

Available Media

In-process system call interposers are increasingly used to extend or monitor application functionality without context-switching or inter-process communication (IPC) overhead. However, when embedded inside untrusted applications, existing solutions face a trade-off: they either enforce no isolation at all, or impose severe restrictions on the monitor's isolated programming environment that are incompatible with complex, real-world use cases.

This paper presents Secpoline, a new interposition platform that allows in-process monitors to support arbitrary interposer functionality without compromising isolation. Secpoline achieves this via isolated multi-program loading and a novel meta-monitor design that interposes the monitor itself to emulate a seamless programming environment. In addition, Secpoline implements the first fully self-contained in-process sandbox to harden its own isolation primitive, while preserving fast-path syscall interposition with zero kernel involvement.

Our evaluation shows that Secpoline matches the efficiency of state-of-the-art secure in-process interposers while enabling a significantly more expressive programming environment. We specifically demonstrate this by implementing a kernel-module-free version of the Falco intrusion detection engine. We also implement an in-process ProxySQL sidecar, embedded directly into the application process, where Secpoline transparently provides a kernel-bypass communication path to increase throughput by roughly 50%. These results confirm that Secpoline finally delivers on the promise of secure, complex application monitoring at in-process speeds.

SoK: Capability Operating Systems: Is the Future Finally Here?

Noah Mauthe, Eric Ackermann, and Sven Bugiel, CISPA Helmholtz Center for Information Security

Available Media

Capability operating systems have existed for six decades, yet we do not have a systematic way to compare them and their different design aspects. We even lack a consensus on the purpose of capabilities in these systems. Given the recent resurgence of interest in capability systems in different domains, we provide a definition of such systems and a set of characteristics to describe and distinguish them. Applying our characteristics to seminal capability systems, we then provide a taxonomy of fundamental design approaches that emerged over the last sixty years. Our results lead us to discuss further research in this field and highlight open challenges in creating pure object-capability systems.

Enjoy the Free Lunch, Someone Paid for Us: Escaping Resource Limits of MicroVM-based Containers

Shiwen Wang, State Key Laboratory of Cyberspace Security Defense, Institute of Information Engineering, CAS, and School of Cyber Security, University of Chinese Academy of Sciences; Wu Luo, State Key Laboratory of Cyberspace Security Defense, Institute of Information Engineering, CAS; Kaicheng Liu and Zheyuan Xu, State Key Laboratory of Cyberspace Security Defense, Institute of Information Engineering, CAS, and School of Cyber Security, University of Chinese Academy of Sciences; Yaowen Zheng, Wenhao Wang, Shijun Zhao, Peinan Li, and Rui Hou, State Key Laboratory of Cyberspace Security Defense, Institute of Information Engineering, CAS

Available Media

MicroVM-based containers are increasingly deployed in public clouds (e.g., AWS, Azure, and Alibaba Cloud) to combine container efficiency with strong isolation. However, we demonstrate that microVM-based architectures introduce a new class of resource accounting vulnerabilities which cause discrepancies between the physical resources consumed by a guest and the resource usage accounted by the host. To systematically uncover such issues, we propose a resource accounting analysis framework that examines accounting inconsistencies across both memory and storage subsystems. Using this framework, we identify three easily exploitable attack vectors: (i) writable memory accounting evasion via shared file-backed mappings, (ii) read-only memory accounting evasion via private file-backed mappings, and (iii) a storage accounting evasion that enables unbounded disk consumption through unmonitored file paths.

We further combine these attack vectors into FREE (File-based Resource Expansion & Exploitation), a consolidated attack that freely inflates both memory and storage usage while evading associated resource accounting and billings. We also present a practical attack implementation capable of seamless integration into real-world applications, leveraging a custom dynamic shared library that transparently redirects standard allocations (e.g., malloc) to unaccounted file-backed mappings. We validate the feasibility of the FREE attack on MicroVM-based containers deployed in AWS, Azure and Alibaba Cloud, and use a billing evaluation model to quantify the economic profit of adversarial tenants. Our evaluation demonstrates that FREE can reduce the attacker's billed cost to 42.12% (Kata) and 41.93% (Firecracker) of the original payment. FREE can also be used to degrade the performance of co-located tenants, including up to 57.1% reduction in memory bandwidth and 67.06% loss in I/O throughput.

Plaintext Recovery Against Post-Filtering Access Control

Zachary Espiritu, MongoDB Research; David Cash, University of Chicago

Available Media

Fine-grained access control (FGAC) mechanisms such as row-level security (RLS) and document-level security (DLS) are widely deployed in databases to restrict access to data stored in physical indexing structures shared by multiple users (e.g., in multi-tenant databases, or in the implementation of least-privilege within an organization). FGAC implementations often use post-filtering where untrusted queries run over all data and private results are redacted afterwards. Prior work shows this approach can lead to side-channels that enable attackers to test if a chosen value exists in unseen data. While damaging, prior attacks do not enable the efficient recovery of rich, high-entropy data like full records or text documents.

We show these side-channels are more damaging than previously thought. Using rich query interfaces (e.g., range, prefix, and conjunctive predicates), we amplify existence leakage into reconstruction attacks. We do this in two settings:

  • PostgreSQL (RLS timing). We exploit a timing side-channel and expressive SQL queries (e.g., ranges, conjunctions) to enumerate unknown attribute values and, in turn, full records via binary search over large domains.
  • Elasticsearch/OpenSearch (DLS scoring). We exploit scoring and prefix-expansion side-channels to recover indexed terms from documents. In some cases, we can extract n-grams in the corpus to recover approximate text.

Our results show that FGAC side-channels must be evaluated in the presence of rich predicates, which can turn membership tests into scalable reconstruction of high-entropy records.

PICASSO: Scaling CHERI Use-After-Free Protection to Millions of Allocations using Colored Capabilities

Merve Gülmez, Ericsson Security Research; Ruben Sturm, DistriNet, KU Leuven; Hossam ElAtali, University of Waterloo; Håkan Englund, Ericsson Security Research; Jonathan Woodruff, University of Cambridge; N. Asokan, University of Waterloo and KTH Royal Institute of Technology; Thomas Nyman, Ericsson Product Security

Available Media

While the CHERI instruction-set architecture extensions for capabilities enable strong spatial memory safety, CHERI lacks built-in temporal safety, particularly for heap allocations. Prior attempts to augment CHERI with temporal safety fall short in terms of scalability, memory overhead, and incomplete security guarantees due to periodical sweeps of the system's memory to individually revoke stale capabilities.

We address these limitations by introducing colored capabilities that add a controlled form of indirection to CHERI's capability model. This enables provenance tracking of capabilities to their respective allocations via a hardware-managed provenance-validity table, allowing bulk retraction of dangling pointers without needing to quarantine freed memory. Colored capabilities significantly reduce the frequency of capability revocation sweeps while improving security.

We realize colored capabilities in PICASSO, an extension of the CHERI-RISC-V architecture on a speculative out-of-order FPGA softcore (CHERI-Toooba). We also integrate colored-capability support into the CheriBSD OS and CHERI-enabled Clang/LLVM toolchain. Our evaluation shows effective mitigation of use-after-free and double-free bugs across all heap-based temporal memory-safety vulnerabilities in NIST Juliet test cases, real-world CVEs, only a small performance overhead on SPEC CPU benchmarks (≈ 5% g.m.), less latency, and more consistent performance in long-running SQLite, PostgreSQL, and gRPC workloads compared to prior work.

Trust Nothing: RTOS Security without Run-Time Software TCB

Eric Ackermann and Sven Bugiel, CISPA Helmholtz Center for Information Security

Available Media

Embedded devices face an ever-expanding threat landscape: vulnerabilities in application software, operating system kernels, and peripherals threaten the embedded device integrity. Existing computer-architectural defenses fully consider at most two of these threat vectors in their security model.

This paper aims at addressing this gap using a novel capability architecture. To this end, we combine a token capability approach suitable for building an untrusted operating system with protection against malicious devices without requiring hardware changes to peripherals.

First, we develop and evaluate a full FPGA implementation of our capability architecture around legacy hardware components. Further, we present a soft real-time operating system based on Zephyr that has no run-time software TCB. To this end, we disaggregate Zephyr's subsystems into small, mutually isolated components. All subsystems that exist at run time, including scheduler, allocator and DMA drivers, and all peripherals are fully untrusted. We believe that our work offers a foundation for more rigorous security-by-design in tomorrow's security-critical embedded devices.

Track 5

Enigma Track: Cryptography

Session Chair: Nick Sullivan, Cryptography Consulting LLC

Dover Room

How Can We Have Private Access to Intelligence?

Tanya Verma, Tinfoil

Available Media

Our interactions with AI models are getting more intimate. It started out with chat, where we voluntarily shared specifics. Now agents are always on and always listening. Today, all context flows through cloud servers in plaintext, accessible to model providers, hackers, and governments. Digital privacy now effectively means protecting your brain.

This talk examines the paths to private AI. Fully homomorphic encryption offers mathematical elegance but remains impractical for frontier-scale inference. However, secure hardware primitives are practical today and will take us very far in accomplishing the goal of building a verifiably private infrastructure for AI. Based on lessons from deploying confidential inference serving billions of tokens daily, I'll chart why this matters now, and how secure hardware, combined with transparency logs, and open-source verification, can make private AI the default.

Tanya is a co-founder of Tinfoil and was previously a Cloudflare engineer and researcher. Tanya designed and deployed privacy and security protocols used by billions of users, and contributed to Cloudflare's Workers AI platform.

Cryptography at Scale: Building a Practical Cryptographic Inventory for Risk, Compliance, and Post‑Quantum Readiness

Jayati Dev, Comcast Cable

Available Media

Large organizations depend on cryptography for security and compliance with existing standards yet often lack a reliable understanding of where cryptographic assets actually exist. "Assets", like, certificates, keys, algorithms, and protocols are scattered across source code, network traffic, infrastructure, and third‑party systems, frequently managed by different teams. This lack of visibility makes it difficult to assess cryptographic risk and create a well-informed plan to migrate to newer quantum-resistant algorithms. Post-quantum readiness largely depends on what we can observe – we cannot fix cryptography that we cannot see. This talk presents XIPHER, a scalable approach to cryptographic inventory. We discuss XIPHER as a methodology and a suite of tools that can build scalable inventory of cryptographic assets to help with current and future compliance with encryption standards.

Jayati Dev is a cybersecurity researcher working at the intersection of policy and emerging technologies at Comcast. She leads the inventorying workstream for the Post-Quantum Cryptography (PQC) Center of Excellence, building tools and solutions for encryption compliance. She also supports the public policy workstream, working on various areas of emerging tech policy like PQC. Jayati holds a PhD in Security Informatics from Indiana University Bloomington where she worked on privacy-preserving technologies in conversational platforms. She has also been a Google Public Policy fellow in cybersecurity policy and a co-lead researcher in a National Science Foundation multi-year investigation into IoT privacy.

What Is a PUF and Why Might You Need One?

Shannon Egan, Deep Science Ventures

Available Media

How does a chip prove its identity? As networked IoT devices and AI clusters increasingly handle and transmit sensitive data, strong device authentication has become more central to commodity hardware design. Physical Unclonable Functions (PUFs) are one way of providing unique device identifiers, leveraging natural variations in the manufacturing process to create a "silicon fingerprint." By tracing PUFs from their origins in smart card security to their emerging role in data centre silicon, I will explain what PUFs uniquely offer and why these properties matter: from device physics, to attacks on OTP memory and implementation challenges.

Shannon Egan is a physicist and multi-disciplinary researcher whose peer-reviewed work spans neuroscience, particle physics, quantum materials, and spintronic device applications in computing. Her research on entropy generation in emerging memory technologies drew her attention to the physical properties underpinning hardware security primitives and secure memory systems. She now leads R&D and product definition in these areas for an early-stage company incubated at Deep Science Ventures.

3:30 pm–4:00 pm

Coffee and Tea Break

Grand Foyer North and West

4:00 pm–5:30 pm

Track 1

Human Aspects 2

Session Chair: Shuang Hao, University of Texas at Dallas

Grand Ballroom VII-X

Security and Privacy Considerations for Confirming Payments in Rwandan Mobile Money Systems

Oluwole A. Adewusi, Assane Gueye, Wallace S. Msagusa, Jema David Ndibwile, David Nkundineza, Okemawo Obadofin, and Wilson Rutaremara, Carnegie Mellon University Africa; Blase Ur, University of Chicago

Distinguished Paper Award Winner

Available Media

Mobile money (MoMo) services let users pay individual merchants or businesses using either a smartphone or a basic phone. They are widely used for commerce in African countries, including Rwanda. To better understand the security and privacy aspects of how Rwandan businesses confirm MoMo payments in practice, we interviewed 30 Rwandan business owners and employees. While MoMo services send payment confirmations over SMS to both senders (customers) and recipients (merchants), we found that merchants often inspect confirmation messages on customers' phones, not their own. The reasons included confirmation messages being sent only to business owners' (not employees') phones, busy markets, phones being physically unavailable, and delayed SMS delivery. Trust also played a major role. As the customer controls their phone, this workaround can enable fraud via screenshots of past messages or spoofed confirmations. As these confirmation messages include customers' account balances, participants also reported privacy concerns. In response, we investigated the design space of alternative workflows, finding promise in redesigning customers' confirmation screens to remove private information and to add a secret known only to the merchant. We reflect on designing for security and privacy under the unique constraints of Sub-Saharan Africa.

Cross-National Information Attacks: A Two-Decade Analysis of Troll Behavior in Korea

Jaehong Kim, Korea Advanced Institute of Science and Technology (KAIST) and Max Planck Institute for Security and Privacy (MPI-SP); Hyeonseung Kim, Korea Advanced Institute of Science and Technology (KAIST); Jiseon Kim, Korea Advanced Institute of Science and Technology (KAIST) and Max Planck Institute for Security and Privacy (MPI-SP); Alice Oh, Korea Advanced Institute of Science and Technology (KAIST); Thorsten Holz, Max Planck Institute for Security and Privacy (MPI-SP); Wonjae Lee, Korea Advanced Institute of Science and Technology (KAIST); Meeyoung Cha, Max Planck Institute for Security and Privacy (MPI-SP) and Korea Advanced Institute of Science and Technology (KAIST)

Available Media

Coordinated foreign influence operations pose a growing threat to online platforms, but detecting state-linked troll activity and tracking its evolution remain challenging. This paper presents an explainable machine learning framework for theory-guided detection and longitudinal analysis of suspected trolling within Korean online news comment sections. Our hierarchical model classifies comments along three dimensions central to influence campaigns: foreign origin, moral-emotional framing, and target country. To support explainability, it also extracts brief span-level textual evidence that provides human-interpretable rationales. We apply the approach to 112M South Korean news comments authored by 4M users over nearly 20 years, identifying 23,998 accounts exhibiting behavior consistent with coordinated manipulation. Analyzing these accounts, we find that they predominantly rely on morally condemning rhetoric rather than direct promotion of foreign-aligned narratives; this rhetoric receives significantly higher user engagement. Among the highest-engagement comments, the moral condemnation most frequently targets domestic political figures (e.g., presidents or party leaders) on both the left and the right, potentially amplifying polarization. Our framework supports transparent platform governance through explainable, evidence-based moderation. These observed rhetorical and engagement patterns can inform how platforms and observatories prioritize defenses and intervene before harmful narrative-target combinations achieve widespread reach.

Exploring Privacy Negotiation Strategies for Camera-Equipped Smart Home Devices in Airbnb

Zixin Wang, University of Michigan, Ann Arbor; Sunyup Park, University of Maryland, College Park; Haojian Jin, University of California, San Diego; Yaxing Yao, Johns Hopkins University

Available Media

Privacy conflicts among multiple users are prevalent in smart environments. To address such conflicts and mitigate the privacy concerns of bystanders, privacy negotiation has been proposed as a promising approach. Yet it remains unclear what makes such negotiations accessible to tenants. In this paper, using camera-equipped smart devices in Airbnb as a case, we explored different privacy negotiation strategies to shed light on future designs to support such actions. We developed 24 storyboards varying four factors, including device types (e.g., driveway camera, video conferencing device for TVs, smart display device), negotiation initiators (host vs. tenant), strategies (physical adjustment vs. monetary compensation), and outcomes (accepted vs. declined). Then, using a vignette survey study (N = 400), we find that device type and strategy significantly shape evaluations: physical adjustment was consistently preferred over monetary compensation, and driveway camera negotiations were evaluated more favorably than those involving indoor devices (video conferencing device for TVs, smart display device). We summarize design implications to support privacy negotiations.

Selling the Dream: How Intimate Insiders and Identity-Based Attackers Disrupt Micro-businesses

Nazanin Sabri, University of California San Diego; Arkaprabha Bhattacharya and Sterling Williams-Ceci, Cornell University; Daniel V. Bailey, Independent Researcher; Rosanna Bellini, New York University

Available Media

Micro-businesses represent the majority of all businesses, and act as vehicles to economic freedom for many at-risk population groups. At the risk of falling behind, many microbusiness owners (MBOs) increasingly adopt platforms as their primary infrastructure, and rely heavily on recruiting friends, families, and even romantic partners for running essential operations. Despite these facts, most prior work has focused on larger corporations and enterprise cybersecurity hygiene, ignoring the cybersecurity risks that such social relationships and changes in visibility introduce.

In this work, we conduct a deep dive into the experiences of MBOs through two complementary analyses: an evaluation of case records of 38 at-risk MBOs subject to insider threats, and semi-structured interviews with 16 MBOs. Our analysis show a combination of restricted access to financial resources and disruption to relationships drove them to take greater risks in recruiting and access control, leaving them more vulnerable to malicious insider attacks. They also shared that their pursuit of market reach frequently led to an unwanted, hypervisible presence, inadvertently providing intimate insiders among other adversaries with further leverage or information needed to cause harm. We conclude by proposing avenues for eCommerce governance that address these risks, focusing on preventing insider harm within these highly visible environments.

WarrInt: Integrity Validation for Criminal Legal Process

Alisha Ukani, UC San Diego; Ross Greer, UC Merced; Katherine Izhikevich, Earlence Fernandes, Alex C. Snoeren, and Stefan Savage, UC San Diego

Available Media

In the U.S., criminal legal process (e.g., a search warrant) is the key mechanism by which law enforcement entities and courts compel third parties (e.g., Google or Meta) to produce evidence in support of a criminal investigation. However, these legal orders are simply documents, commonly served electronically, and there is no inherent mechanism to guarantee their integrity. In recent years, law enforcement has increasingly reported the use of forged legal process used to advance a variety of malign purposes, with online sellers offering such capabilities for a fee. We design a practical system, WarrInt, that is compatible with the constraints of existing process mechanics and helps third parties identify when the legal orders they receive may be forged or tampered with. Our evaluation on real and modified legal process documents shows that WarrInt is effective at identifying modifications.

Characterizing Security and Privacy Teaching Standards for Schools in the United States

Katherine Limes, Colorado School of Mines; Nathan Malkin, New Jersey Institute of Technology; Kelsey R. Fulton, Colorado School of Mines

Available Media

Increasingly, students begin learning aspects of security and privacy during their primary and secondary education (grades K-12 in the United States). Individual U.S. states and some national organizations publish teaching standards—guidance that outlines expectations for what students should learn. However, research has not yet examined what is covered by these standards and whether the topics align with what the broader security and privacy community thinks students should know. To shed light on these questions, we manually examined 12,201 standards from all U.S. states and eight national organizations. We labeled 3,920 of them as being related to security and privacy, further classifying these into 109 topics ranging from encryption to appropriate online behavior.

A Midsummer Meme's Dream: Investigating Market Manipulations in the Meme Coin Ecosystem

Alberto Maria Mongardini, Sapienza University of Rome and Technical University of Denmark; Alessandro Mei, Sapienza University of Rome

Available Media

From viral jokes to a billion-dollar phenomenon, meme coins have become one of the most popular segments in cryptocurrency markets. Unlike utility-focused crypto assets like Bitcoin, meme coins derive value primarily from community sentiment, making them vulnerable to manipulation. This study presents an unprecedented cross-chain analysis of the meme coin ecosystem, examining 34,988 tokens across Ethereum, BNB Smart Chain, Solana, and Base. We characterize their tokenomics and track their growth in a three-month longitudinal analysis. We discover that among high-return tokens (>100%), an alarming 82.89% show evidence of artificial growth strategies designed to create a misleading appearance of market interest. These include wash trading and a new form of manipulation we define as Liquidity Pool-Based Price Inflation (LPI), where small strategic purchases trigger dramatic price increases. We find that profit extraction schemes, such as pump and dumps and rug pulls, typically follow initial manipulations like wash trading or LPI, indicating how early manipulations create the foundation for later exploitation. We quantify the economic impact of these schemes, identifying over 17,000 victim addresses with realized losses exceeding $9.3 million. These findings reveal that combined manipulations are widespread among high-performing meme coins, suggesting that their dramatic gains are often driven by coordinated efforts rather than natural market dynamics.

Resilience Amid Diffused Surveillance During a Political Movement in Bangladesh

Mashiyat Mahjabin Eshita, BRAC University; Ishmam Bin Rofi, Rochester Institute of Technology; Md. Sabbir Ahmed, BRAC University; Dipto Das, University of Toronto; Md. Golam Rabiul Alam, BRAC University; Jannatun Noor, C2SG Research Group, United International University; Syed Ishtiaque Ahmed, University of Toronto; S M Taiabul Haque, BRAC University

Available Media

In July 2024, Bangladesh witnessed a historic, student-led uprising where social media became both a tool of mobilization and a site of repression. While surveillance is often framed through Bentham's vertical Panopticon model, our study extends the discourse on lateral surveillance. Drawing on 23 interviews with protesters, activists, and digital organizers, we examine how lateral surveillance operates within a political movement in Bangladesh, where discipline is enacted not only by the state but also through peers, kin, and intimate social ties. Participants described living under constant fear while also developing creative strategies of resilience, from improvised encrypted systems and panic deletions to cultural tactics like satire and hidden speech. We contribute to security and privacy literature on activism by highlighting the need for designs that strengthen safety, anonymity, and collective resilience in fragile and repressive contexts.

Information Security in Small-Scale Protests: Surveillance of Ugandan Anti-EACOP Protesters

Ntezi Mbabazi and Rikke Bjerg Jensen, Royal Holloway, University of London

Distinguished Paper Award Winner

Available Media

We examine the information security practices of Ugandan climate activists protesting the development of the East African Crude Oil Pipeline (EACOP). We conducted five-week fieldwork in Kampala, Uganda, which included interviews with 13 anti-EACOP activists. Through an inductive analysis, we report on the complexities faced by small groups of predominantly student protesters as they covertly organise small-scale anti-EACOP protests within a context marked by state surveillance and repression. Our study points to a multi-layered adversarial landscape, where participants' experiences of direct threats, including arrests and information compromise, and their fears of abduction, shaped their security practices. These practices were rooted in autonomous decision-making within groups. We present a grounded understanding of how participants' need to protect information for their own security, as well as that of others, permeated their lives, leading them to adjust day-to-day aspects of their device management, communication, accommodation, transport and social relations as deliberate tactics to mitigate surveillance.

Track 2

Software Security 3

Session Chair: Qinying Wang, EPFL

Grand Ballroom VI

Autonomy Comes with Costs: Detecting Denial-of-Service Vulnerabilities Caused by Resource Abusing in LLM-based Agents

Jiaqi Luo, Jiarun Dai, Fengyu Liu, Songyang Peng, Youkun Shi, Tong Bu, and Geng Hong, Fudan University; Xudong Pan, Fudan University and Shanghai Innovation Institute; Yuan Zhang, Fudan University

Available Media

LLM-based agents have recently attracted significant attention. By leveraging the semantic understanding capabilities of large language models (LLMs), these agents can autonomously perform complex tasks according to user requests, such as downloading files and summarizing content. However, the lack of comprehensive resource governance renders them susceptible to abuse, potentially leading to resource exhaustion and denial-of-service (DoS) conditions.

In this work, we present the first systematic security study of resource management in LLM-based agents. We identify three representative patterns of resource lifecycle management, each of which enables distinct avenues for DoS exploitation. Building on these insights, we propose AgentDoS, a novel directed grey-box fuzzing framework designed to detect DoS vulnerabilities arising from resource exhaustion. AgentDoS first analyzes the resource lifecycle within the agent and then leverages an LLM to generate functionality-specific seed prompts in natural language that drive the agent toward excessive resource consumption. We evaluated AgentDoS on 20 widely used open-source LLM-based agents and discovered 36 zero-day vulnerabilities affecting 16 agents, 15 of which have over 10,000 stars on GitHub. To date, 15 CVE IDs have been assigned for these vulnerabilities.

BugAuditor: Detecting Bugs via Inconsistent Defensive Code Auditing

Miaoqian Lin, The University of Hong Kong; Kai Chen, Institute of Information Engineering, Chinese Academy of Sciences; Hao Chen, The University of Hong Kong

Available Media

Modern software systems contain complex behaviors that are prone to bugs when handled incorrectly. Detecting such bugs requires reliable oracles, which are difficult to construct as they require project-specific knowledge. Prior studies mainly obtain oracles from comparable code deviations, documentation, or historical bugs. However, these sources are inherently limited, leaving many project-specific bugs undetected.

In this paper, we propose inconsistent defensive handling as a new bug oracle for LLM-driven bug auditing. Specifically, we observe that real-world systems contain abundant defensive code, where developers proactively apply defensive handling in security-sensitive contexts to prevent bugs, and inconsistencies in such handling can indicate bugs. To realize this idea, we design BugAuditor, a framework that performs bug auditing via inconsistent defensive handling. BugAuditor first locates defensive code snippets across the codebase. It then reasons about the code to infer its security intent and underlying defensive patterns, characterizing the associated security-sensitive behaviors and defensive handling. Finally, BugAuditor applies the inferred patterns to audit the codebase and reports inconsistent defensive handling of the same security-sensitive behaviors across different contexts.

We evaluate BugAuditor on the Linux kernel. The results show that BugAuditor effectively mines project-specific knowledge embedded in defensive code that existing methods miss. Using the inferred defensive patterns, BugAuditor detects 54 long-latent bugs, including resource leaks, information leaks, and invalid pointer dereferences. To date, 20 bugs have been confirmed and fixed in the latest version, and two have been assigned CVE identifiers.

TrioFuzz: A Three-Tier Architecture for Adaptive Strategy Selection in Fuzzing

Ruiqi Dong, Swinburne University of Technology; Yiyi Wang, Tsinghua University and Zhongguancun Laboratory; Kunpeng Zhang, The Hong Kong University of Science and Technology; Dongsong Yu, Zhongguancun Laboratory; Xiaogang Zhu, Adelaide University; Shaohua Wang, Central University of Finance and Economics; Shuai Wang, The Hong Kong University of Science and Technology; Chao Zhang, Tsinghua University and Zhongguancun Laboratory; Sheng Wen and Yang Xiang, Swinburne University of Technology

Available Media

Adaptive strategy selection is a promising direction for improving fuzzing effectiveness, yet existing learning-based approaches often fail to outperform random selection in practice. We identify the root cause as an architectural mismatch rather than an algorithmic limitation. Our empirical study reveals that strategy effectiveness shifts at minute-level timescales, while existing in-loop architectures require tens of minutes per learning cycle. This order-of-magnitude gap prevents timely adaptation regardless of which learning algorithm is used.

We propose TrioFuzz, a new fuzzing engine that decouples learning from execution through a three-tier thread architecture. A centralized learning thread aggregates feedback from parallel execution threads, compressing the adaptation cycle to under one minute. TrioFuzz integrates the same learning algorithms as prior work, isolating the architectural contribution from algorithmic factors. Our ablation study shows that learning algorithms perform significantly better under our proposed architecture than under their original in-loop implementations. Notably, native MOpt underperforms baselines, but becomes the top performer under TrioFuzz's architecture. Evaluation on FuzzBench and OSS-Fuzz shows that TrioFuzz achieves the highest coverage on 7 out of 9 FuzzBench targets and 8 out of 10 OSS-Fuzz projects, triggers 17% more vulnerabilities than AFL++ on Magma, and discovers 19 new CVEs in 6 real-world projects

FirmReBugger: A Benchmark Framework for Monolithic Firmware Fuzzers

Mathew Duong, Michael Chesser, and Guy Farrelly, University of Adelaide; Surya Nepal, Data61 CSIRO; Damith C. Ranasinghe, University of Adelaide

Available Media

Monolithic Firmware is widespread. Unsurprisingly, fuzz testing firmware is an active research field with new advances addressing the unique challenges in the domain. However, understanding and evaluating improvements by deriving metrics such as code coverage and unique crashes are problematic, leading to a desire for a reliable bug-based benchmark. To address the need, we design and build FirmReBugger, a holistic framework for fairly assessing monolithic firmware fuzzers with a realistic, diverse, bug-based benchmark.

FirmReBugger proposes using bug oracles—C syntax expressions of bug descriptors—with an interpreter to automate analysis and accurately report on bugs discovered, discriminating between states of detected, triggered, reached and not reached. Importantly, our idea of benchmarking does not modify the target binary and simply replays fuzzing seeds to isolate the benchmark implementation from the fuzzer while providing a simple means to extend with new bug oracles.

Further, analyzing fuzzing roadblocks, we created FirmBench, a set of diverse, real-world binary targets with 313 software bug oracles. Incorporating our analysis of roadblocks challenging monolithic firmware fuzzing, the bench provides for rapid evaluation of future advances. We implement FirmReBugger in a FuzzBench-for-Firmware type service and use FirmBench to evaluate 9 state-of-the art monolithic firmware fuzzers in the style of a reproducibility study, using a 10 CPU-year effort, to report our findings.

Generating Precise Format Specification for Network Protocols Through Adversarial LLM Interactions

Hengdi Ye, Bing Shui, Jielun Wu, Yufan Zhou, Baowen Xu, and Qingkai Shi, State Key Laboratory of Novel Software Technology, Nanjing University

Available Media

This paper aims to address a significant gap in inferring precise format specifications for network protocols, particularly regarding the exact constraints on network packet fields. That is, while current methods can often derive accurate syntactic structures that divide a network packet into multiple fields, they typically struggle to capture the semantic dependencies or constraints among those fields. To address this issue, our central insight is that large language models (LLMs) have proven effective across various tasks, and adversarial interactions among these tasks can significantly alleviate the hallucination problem. Based on the insight, we propose a novel approach to generating precise protocol formats by adversarially combining LLM-based specification inference and code generation. By inputting the structured RFC documents of network protocols into LLMs, we generate both packet formats and reference packet parsers that iteratively refine one another to reduce hallucinations: the packet formats allow us to create a variety of network packets for parser testing, while runtime checks in the parsers help identify format errors. Our evaluation of 8 protocols shows a significant increase in precision and recall, achieving a 326% improvement over the state of the art in inferring field constraints. Furthermore, the improved format specifications enable effective fuzzing, leading to the discovery of 24 zero-day vulnerabilities in widely used protocol implementations.

Bulbasaur: Branch-Guided Online Mutator Generation for Greybox Fuzzing

Yiyi Wang, Tsinghua University and Zhongguancun Laboratory; Dongsong Yu, Zhongguancun Laboratory; Ruiqi Dong, Swinburne University of Technology; Yiyang Chen, Tsinghua University; Xiaogang Zhu, Adelaide University; Chao Zhang, Tsinghua University and Zhongguancun Laboratory

Available Media

Coverage-guided greybox fuzzing (CGF) hinges on generating inputs that satisfy branch constraints in order to explore deep code regions. Existing CGF techniques propose various approaches to improve the probability of solving such constraints. However, they still rely on general-purpose mutators, which lack the ability to precisely manipulate relevant input bytes to satisfy complex constraints.

Our analysis reveals that LLMs can more accurately solve branch constraints by generating tailored mutators. These mutators enable the fuzzer to perform localized and reusable modifications on inputs that reach target branches, producing precise mutations to satisfy target constraints. To achieve this, we propose Bulbasaur, a branch-guided framework for online LLM-based mutator generation. Bulbasaur employs hard frontier-guided branch selection to identify critical branches, continuously collects and organizes relevant static and dynamic context to support high-quality mutator generation, and adopts an efficient and adaptive strategy to apply generated mutators during fuzzing.

We evaluate Bulbasaur on the FuzzBench and UniFuzz benchmarks. The results show that Bulbasaur improves line and branch coverage by up to 23.18% and 24.46%, respectively, over state-of-the-art fuzzers, discovers 21 vulnerabilities missed by all baselines, and identifies 15 previously unknown vulnerabilities on OSS-Fuzz, all of which have been assigned CVE identifiers.

Death by a Thousand Drips: Uncovering Critical Resource Leaks in the Windows Ecosystem

Feng Dong and Jianting Gao, Huazhong University of Science and Technology; Yunpeng Tian, The Hong Kong Polytechnic University; Weifeng Yuan, Huazhong University of Science and Technology; Mu Zhang, University of Utah; Zesen Ye and Jietao Yang, Beijing CyberKunlun Technology Co., Ltd; Zhiniang Peng, Huazhong University of Science and Technology

Available Media

Windows remote services are a critical attack surface due to their privileged execution and widespread deployment. Under the shared service-host model (svchost.exe), a single resource-exhaustion bug can trigger fate-sharing failures across co-resident services. Yet most vulnerability discovery remains crash-centric and is thus blind to silent, long-horizon resource leaks (e.g., memory/handle exhaustion) that do not cause immediate exceptions, while the closed-source Windows ecosystem limits the applicability of source-level sanitizers. We present REDOSPECTOR, a fuzzing framework for discovering resource-oriented Denial-of-Service (DoS) vulnerabilities in opaque Windows binaries. REDOSPECTOR combines (1) lightweight in-process hook-based telemetry by intercepting native ntdll heap routines, (2) an amplification-based oracle that identifies persistent leaks via repeated execution and growth-trend analysis, and (3) a hybrid seed-generation pipeline that leverages static binary analysis and Large Language Models (LLMs) to reach deep protocol states. Evaluated on the latest builds of Windows 11 (24H2) and Windows Server Preview, REDOSPECTOR uncovers 19 vulnerabilities in core services (including MSMQ, Remote Desktop Gateway, and CDP), leading to 12 assigned CVEs. Microsoft confirmed seven pre-authentication DoS issues in CDP and performed architectural adjustments, demonstrating that REDOSPECTOR can expose resource-mismanagement flaws that evade conventional fuzzing.

RecStruct: Recovering Nested Struct Types from Stripped Binaries via Stack-Driven Unification

Yuxin Chen, Zhiyang Fang, Shiyi Wu, Yixin Xu, and Yuhang Wang, Sichuan University; Xiaokang Yin, Information Engineering University; Junfeng Wang, Sichuan University

Available Media

Recovering composite data types from stripped binaries is critical for reverse engineering and vulnerability analysis. Existing approaches either rely on heavyweight decompilers with limited type reasoning, or employ expensive constraint solvers that fail to scale. While some methods can handle nested structures, they struggle with large real-world binaries due to prohibitive computational costs. We present RecStruct, a fast type recovery framework that reconstructs nested structures without dependence on any decompiler. RecStruct lifts binaries directly, performs flow-insensitive symbolic execution to reduce memory aliases, and introduces Stack-Driven Unification—a novel union-find algorithm that maintains an offset stack during traversal to merge types only when their nesting depths align. This enables precise recovery of recursive structures. We implement an end-to-end pipeline and evaluate it on 418 binaries comprising 744,770 functions, including benchmarks from prior work and large-scale programs such as QEMU and OpenSSL. Experimental results demonstrate competitive accuracy compared to existing tools, with substantial improvements in efficiency and scalability, enabling analysis of binaries that were previously intractable.

CombiSan: Unifying Software Sanitizers for Comprehensive Fuzzing

Matteo Marini, Sapienza University of Rome; Floris Gorter, Vrije Universiteit Amsterdam; Daniele Cono D'Elia, Sapienza University of Rome; Cristiano Giuffrida, Vrije Universiteit Amsterdam

Available Media

Modern C/C++ bug detection efforts heavily rely on fuzzing with software sanitizers. However, the most popular sanitizers have limited interoperability. As a result, developers often enable each sanitizer in isolation, if at all, requiring multiple runs. This sequential execution undermines performance and tests code in a non-uniform manner.

In this paper, we present CombiSan, a fuzzing-optimized sanitizer that simultaneously detects all the addressability, uninitialized memory, and other undefined behavior issues covered by the three most popular sanitizers: ASan, MSan, and UBSan. CombiSan features a unified shadow memory design that efficiently tracks both the addressability and the initialization state of every byte of program memory. In addition, CombiSan's instrumentation seamlessly integrates with state-of-the-art detection of other undefined behavior classes. As bugs found by different sanitizers may mask each other by terminating execution early, CombiSan defers its analysis of all aggregated issues to test case completion.

In our evaluation, CombiSan detected 81 new bugs in 10 programs tested daily by OSS-Fuzz. On average, fuzzing with CombiSan is 1.7x faster than sequentially testing with ASan+UBSan and MSan. Moreover, our results demonstrate that CombiSan has the same bug detection accuracy as these sanitizers, despite running for significantly fewer CPU hours.

Track 3

Applications of Cryptography 4

Session Chair: Eyal Ronen, Tel Aviv University

Grand Ballroom V

End-to-End Encrypted Collaborative Documents

Christian Knabenhans, EPFL; Zayd Maradni, Max Planck Institute for Software Systems; Carmela Troncoso, Max Planck Institute for Security and Privacy & EPFL

Available Media

Collaborative documents (e.g., Google Docs, Microsoft 365) often contain sensitive information such as personal or financial data. In this work, we extend the protection of E2EE encryption, currently (mostly) restricted to the use case of messaging, to collaborative documents. We elicit and formalize the security and functional requirements of End-to-End-Encrypted Collaborative Documents. We then put forth a generic framework to realize E2EE-CD, by combining an end-to-end encrypted asynchronous broadcast channel with any edit reconciliation mechanism which ensures globally consistent views of a document. We give formal proofs that directly relate the security of our E2EE-CD solution to the security of the underlying end-to-end encrypted communication channel. We then elicit additional deployment requirements for E2EE-CD for investigative journalists and design SignalCD, an E2EE-CD system built on top of Signal's group messaging protocol tailored for this setting. We analyze the security guarantees of SignalCD, implement a prototype, and empirically show that our solution is efficient enough to permit real-time collaboration.

Crypto Wars in Secure Messaging: Covert Channels in Signal Despite Leaked Keys

Rosario Giustolisi, IT University of Copenhagen; Gabriele Lenzini, University of Luxembourg; Chuanwei Lin, University of Luxembourg and Karlsruhe Institute of Technology; Mohammadamin Rakeei, University of Luxembourg; Andy Rupp, University of Luxembourg and KASTEL Security Research Labs

Available Media

End-to-end encryption (E2EE) is the foundation of modern secure messaging, with the Signal protocol as the de facto standard in applications such as Signal, WhatsApp, Facebook Messenger and Google Messages. At the same time, the deployment of E2EE has led to growing pressure from authorities to decrypt user traffic under law enforcement. This raises a critical question: if an adversary can routinely decrypt Signal messages (for example via a mandated access or a leaked key), can users still communicate securely and covertly?

We address this question through the lens of anamorphic encryption, which enables hidden communication within seemingly legitimate ciphertexts, even against an adversary who can decrypt them. We design two constructions that embed covert channels into the existing Signal Double Ratchet protocol. Concretely, we show how to embed covert messages (i) into Diffie-Hellman keys used in the asymmetric ratchet, or (ii) into authentication tags produced in the symmetric ratchet. Our techniques are compatible with existing Signal-style deployments and require no changes by the service provider.

We formalize security in threat models that capture adversaries with decryption capabilities granted through lawful-access mechanisms, and prove that the resulting protocol transcripts are indistinguishable from those of standard Signal. We implement our constructions in the official Signal library and Android client, and show that they incur low overhead and are practical in real-world settings. Our results show that covert communication channels can persist even when conventional E2EE guarantees are compromised.

Oblivious Signaling

Mirza Kamrul Bashar Shuhan, George Mason University; Foteini Baldimtsi, George Mason University & Mysten Labs; Giuseppe Ateniese, George Mason University

Available Media

An anonymous messaging service has to solve a basic routing problem: a server must deliver an encrypted message to its recipient without learning who the recipient is. Broadcasting all ciphertexts hides the destination but forces every recipient to constantly scan for new messages. Oblivious Message Retrieval (OMR; CRYPTO~'22) tackles this by using fully homomorphic encryption (FHE) to let an untrusted server perform message retrieval on a recipient's behalf without learning which messages are pertinent.

We introduce Oblivious Signaling, which shifts this cost from retrieval to sending. The server maintains a fixed-size encrypted inbox for each recipient. When a sender submits a message, the server applies the same homomorphic update to every inbox: the intended inbox absorbs the message, and the rest remain unchanged at the plaintext level. The update is uniform, can be parallelized across inboxes, and ties the delivery cost strictly to the size of the anonymity set rather than global traffic. Recipients retrieve by fetching and decrypting their inbox, so checking for new messages is independent of the global traffic.

We formalize receiver privacy against an untrusted server, even when it colludes with other users, give a concrete construction based on fully homomorphic encryption, and analyze the resulting ``digital postage'' trade-off: delivery is expensive, but checking is cheap. Our prototype identifies practical regimes in which this cost-model shift is preferable to scan-based retrieval, even with highly optimized OMR implementations. This cost model is well-suited to settings where recipients check frequently, and messages arrive sporadically, and it naturally discourages high-volume spam.

InstantOMR: Oblivious Message Retrieval with Low Latency and Optimal Parallelizability

Haofei Liang, Shanghai Jiao Tong University; Zeyu Liu, Yale University; Eran Tromer, Boston University; Xiang Xie, Primus Labs; Yu Yu, Shanghai Jiao Tong University

Available Media

Oblivious message retrieval (OMR) addresses the expensive message retrieval process in anonymous messaging systems and private blockchains. It enables resource-limited recipients to outsource detection and retrieval of their messages, while preserving privacy.

This work introduces InstantOMR, a novel OMR scheme that combines TFHE functional bootstrapping with standard RLWE operations in a hybrid design. InstantOMR is specifically optimized for low latency and high parallelizability. Our implementation, using the Primus-fhe library (and estimates based on TFHE-rs), demonstrates that InstantOMR offers the following key advantages:

  • Low latency: InstantOMR achieves \sim 600× lower latency than SophOMR, the state-of-the-art single-server OMR. This translates directly into reduced recipient waiting time (by the same factor) in the streaming setting, where the detector processes incoming messages on-the-fly and returns a digest immediately upon the recipient becoming online.
  • Optimal parallelizability: InstantOMR scales near-optimally with available CPU cores (by processing messages independently), so for high core counts it is faster than SophOMR (whose parallelism is constrained by reliance on BFV).

Message Injection Attacks Against Signal

Kien Tuong Truong, ETH Zurich; Noemi Terzo, Max Planck Institute for Security and Privacy; Kenneth G. Paterson, ETH Zurich

Available Media

Signal is a secure messaging app offering end-to-end security for pairwise and group communications. It has tens of millions of users, and has heavily influenced the design of other secure messaging apps (including WhatsApp). Signal has been heavily analysed and, as a result, is rightly regarded as setting the "gold standard" for messaging apps by the scientific community. We present two practical attacks that break the integrity properties of Signal in its advertised threat model. Each attack arises from different features of Signal that are poorly documented and have eluded formal security analyses. The first attack, affecting Android and Desktop, arises from Signal's introduction of identities based on usernames (instead of phone numbers) in early 2022. We show that the protocol for resolving identities based on usernames and on phone numbers introduced a vulnerability that allows a malicious server to inject arbitrary messages into one-to-one conversations under specific circumstances. The injection causes a user-visible alert about a change of safety numbers, but if the users compare their safety numbers, they will be correct. The second attack is even more severe. It arises from Signal's Sealed Sender (SSS) feature, designed to allow sender identities to be hidden. We show that a combination of two errors in the SSS implementation in Android allows a malicious server to inject arbitrary messages into both one-to-one and group conversations. The errors relate to missing key checks and the loss of context when cryptographic processing is distributed across multiple software components. The attack is undetectable by users and can be mounted at any time, without any preconditions. As far as we can tell, the vulnerability has been present since the introduction of SSS in 2018. We disclosed both attacks to the engineers at Signal who promptly developed and deployed patches. Beyond presenting these devastating attacks on Signal's end-to-end security guarantees, we discuss more broadly what can be learned about the challenges of deploying new security features in complex software projects.

Zero Knowledge (About) Encryption: A Comparative Security Analysis of Three Cloud-based Password Managers

Matteo Scarlata, ETH Zurich; Giovanni Torrisi and Matilda Backendal, USI, Lugano; Kenneth G. Paterson, ETH Zurich

Available Media

Zero Knowledge Encryption is a term widely used by vendors of cloud-based password managers. Although it has no strict technical meaning, the term conveys the idea that the server, who stores encrypted password vaults on behalf of users, is unable to learn anything about the contents of those vaults. The security claims made by vendors imply that this should hold even if the server is fully malicious. This threat model is justified in practice by the high sensitivity of vault data, which makes password manager servers an attractive target for breaches (as evidenced by a history of attacks).

We examine the extent to which security against a fully malicious server holds true for three leading vendors who make the Zero Knowledge Encryption claim: Bitwarden, LastPass and Dashlane. Collectively, they have more than 60 million users and 23% market share. We present 12 distinct attacks against Bitwarden, 7 against LastPass and 6 against Dashlane. The attacks range in severity, from integrity violations of targeted user vaults to the complete compromise of all the vaults associated with an organisation. The majority of the attacks allow recovery of passwords. We have disclosed our findings to the vendors and remediation is underway.

Our attacks showcase the importance of considering the malicious server threat model for cloud-based password managers. Despite vendors' attempts to achieve security in this setting, we uncover several common design anti-patterns and cryptographic misconceptions that resulted in vulnerabilities. We discuss possible mitigations and also reflect more broadly on what can be learned from our analysis by developers of end-to-end encrypted systems.

Cryptanalysis of LDPC-Based Pseudorandom Error-Correcting Codes

Tianrui Wang, Institute for Network Sciences and Cyberspace, BNRist, Tsinghua University; Anyu Wang, Institute for Advanced Study and State Key Laboratory of Cryptography and Digital Economy Security, Tsinghua University; Zhongguancun Laboratory; Tianshuo Cong, School of Cryptologic Science and Engineering and Shandong Key Laboratory of Artificial Intelligence Security, Shandong University; Delong Ran, Institute for Network Sciences and Cyberspace, BNRist, Tsinghua University; Jinyuan Liu, Institute for Advanced Study, Tsinghua University; Xiaoyun Wang, Institute for Advanced Study and State Key Laboratory of Cryptography and Digital Economy Security, Tsinghua University; Zhongguancun Laboratory; National Financial Cryptography Research Center; Shandong Institute of Blockchain

Available Media

Pseudorandom error-correcting codes (PRCs), a novel cryptographic primitive recently proposed at CRYPTO 2024, are primarily applied in undetectable watermarking schemes for large generative models. However, the security of PRCs has not yet been systematically analyzed. To fill this gap, we present the first cryptanalysis of PRCs. Specifically, focusing on LDPC-PRC, the only known practical instantiation of PRCs, we propose three novel attacks that challenge its undetectability and robustness. To rigorously demonstrate the practical threat, we analyze the concrete attack complexity under realistic parameters and validate the attack effectiveness on both real-world large language models and generative image models, including DeepSeek and Stable Diffusion. Our analysis shows that the claimed security guarantees of LDPC-PRC are undermined across all practically feasible regimes. For example, our attacks can detect the presence of a watermark with overwhelming probability at a cost of 222 operations. Beyond attacks, we further propose three defenses: parameter recommendation, implementation suggestion, and a revised key generation function. However, PRC-based watermarking schemes still fail to achieve 128-bit security due to inherent constraints of large generative models, such as the maximum output length of large language models. Overall, our work clarifies the concrete security limits of PRCs in real-world watermarking applications.

Transparent Dictionaries from Polynomial Commitments

Hossein Hafezi, NYU; Alireza Shirzad, University of Pennsylvania; Benedikt Bünz and Joseph Bonneau, NYU

Available Media

We present IRONDICT, a transparent dictionary construction based on polynomial commitment schemes. Transparent dictionaries enable an untrusted server to maintain a mutable dictionary and provably serve clients lookup queries. A major open challenge is supporting efficient auditing by lightweight clients. Previous solutions either incurred high server costs (limiting throughput) or high client lookup verification costs, hindering them from modern messaging key transparency deployments with billions of users. Our construction makes black-box use of a generic multilinear polynomial commitment scheme and inherits its security notions, i.e. binding and zero-knowledge. We implement our construction with the recent KZH scheme and find that a dictionary with 1 billion entries can be verified on a consumer-grade laptop in 35 milliseconds, a 300 -fold improvement over the state of the art. Our construction also offers 150000x smaller proofs ( 8 Kilobytes) and perfect privacy, with concretely efficient client and server costs. We also show fast-forwarding techniques based on incremental verifiable computation (IVC) and checkpoints to enable even faster client auditing.

Anamorphic Messaging: Analyzing the Double Ratchet, Triple Ratchet, PQ3, and MLS

Hien Chu, TU Wien; Alessandro Corsi and Paul Rösler, FAU Erlangen-Nürnberg

Available Media

Anamorphic cryptography targets the scenario in which a dictator does not forbid the use of cryptography but requires all users to reveal their secret keys to them. Thus, the dictator can decrypt all honestly generated ciphertexts. The approach for bypassing this is to identify spots, such as random nonces, in existing cryptographic protocols in which secret messages can be hidden using an additional secret double key. So far, the literature mostly focused on identifying such spots in simple primitives like public-key encryption or signatures; only recently, an initial work identified limited spots in Signal's Double Ratchet Algorithm.

We are the first to leverage the statefulness of cryptographic communication protocols to employ continuously updated double states and, thereby, achieve Forward Security: Even if the adversary (i) observes all traffic, (ii) knows all users' regular secret key material at any stage of the protocol execution, and (iii) at some point learns the secret double state, the entire protocol execution looks benign although covert messages were previously hidden in the traffic anamorphically. We formalize this notion and also cover robustness and authenticity, which appear to be particularly relevant in the messaging context.

In this new model, we study four of the most relevant messaging protocols and identify hiding spots therein: Signal's Double Ratchet, Signal's Triple Ratchet, Apple's PQ3, and the two-party core of the Messaging Layer Security Standard. We focus on the cryptographic parts of these protocols and, despite their complexity, identify surprisingly few anamorphic hiding spots. We prove that all these protocols offer forward secure, authenticated anamorphic channels and we evaluate their bandwidths: While 16 bits can be embedded in every epoch of the Double Ratchet, Triple Ratchet and PQ3 provide 176 bits, respectively 256 bits, of bandwidth per post-quantum epoch, and MLS provides 688 bits per epoch.

Track 4

Security of ML 4

Session Chair: Piotr Mardziel, RealmLabsAI

Grand Ballroom I-IV

ORPHEUS: A Separation-Robust Proactive Defense for Singing Voice Conversion

Zhaolin Wei, School of Cyber Science and Engineering, Wuhan University; Dengpan Ye, School of Cyber Science and Engineering, Wuhan University, and Cyberspace Institute of Advanced Technology, Guangzhou University; Yanjiao Chen, College of Electrical Engineering, Zhejiang University; Jiacheng Deng, Ziyi Liu, Yuhan Lin, and Yunna Lv, School of Cyber Science and Engineering, Wuhan University; Yueyun Shang, School of Mathematics and Information Science, Guangzhou University; Zhihong Tian, Cyberspace Institute of Advanced Technology, Guangzhou University

Available Media

Recent advances in singing voice conversion enable realistic cloning of a singer's voice, raising concerns about unauthorized voice misuse. Existing proactive defenses inject imperceptible perturbations to disrupt such conversion, and they are designed for vocal signals rather than mixed music. In real-world scenarios, however, attackers usually obtain mixed music and apply source separation to decompose the mixture into vocal and backing tracks before conversion. Since separation models are trained to distinguish vocals from background components, perturbations are often treated as backing, causing the perturbations to be filtered out. To address this, we present ORPHEUS, the first proactive defense framework tailored for singing voice conversion involving source separation. Specifically, we jointly (i) interfere with the separation model via mask-misguiding and cross-track losses, (ii) disrupt identity information using an ensemble of heterogeneous speaker encoders to enhance transferability, and (iii) optimize perceptual quality through tonality harmony constraints and psychoacoustic masking. To evaluate ORPHEUS, we conduct experiments on two datasets with three source separation models and four singing voice conversion systems. Compared with the best-performing baseline, ORPHEUS further reduces identity similarity by 8.02% and 5.61% on two speaker verification models, respectively, demonstrating consistently stronger defense effectiveness and cross-model transferability.

On Improving Robustness of Deepfake Image Detectors

Abu Taib Mohammed Shahjahan, Mohammad Mannan, Abdessamad Ben Hamza, and Amr Youssef, Concordia University

Available Media

The rapid advancement of Generative AI has introduced remarkable opportunities while simultaneously raising critical concerns regarding content authenticity. While recent work has increasingly focused on improving the generalization of deepfake detectors across unseen generative models, their robustness against adversarial attacks remains limited. In particular, Abdullah et al. (IEEE SP 2024) evaluated eight detectors and demonstrated that most of them exhibit significant performance degradation under adversarial attacks. We also observed the same phenomenon by testing seven most recent state-of-the-art detectors. To address this problem, we propose a unified framework that integrates three complementary design principles without relying on adversarial training data: (i) higher-order statistical modeling in the frequency domain via Discrete Cosine Transform (DCT)-based moment pooling up to fourth order, (ii) content-agnostic feature representations derived from noise residuals, and (iii) cross-scene generalization enforced through patch-level semantic disruption. A key insight underpinning our approach is that adversarial attacks primarily operate on low-order statistics and visual semantics, leaving higher-order residual-frequency characteristics, particularly kurtosis, largely unconstrained. Extensive experiments demonstrate that our method consistently improves robustness across six architecturally diverse detectors. Notably, we achieve up to 88.9% reduction in recall degradation on current adversarial benchmarks, and improve the best-performing recent detector (Yang et al., IEEE CVPR 2025) from 81.9% to 97.15% accuracy under attack. Overall, our method provides a principled, architecture-agnostic approach for improving deepfake detection robustness against current attacks.

SafeAdapt: Safety Alignment with Adaptive Thinking Allocation for Large Reasoning Models

Jiazheng Song, Zhejiang University; Junxu Liu, The Hong Kong Polytechnic University; Jian Lou, Sun Yat-sen University; Jinfei Liu, Zhejiang University

Available Media

Reasoning models have garnered growing importance as their strong Chain-of-Thought (CoT) capabilities unleash exceptional performance on complex tasks. Consequently, an emerging research direction explores "safety thinking" which leverages reasoning models' reasoning capabilities to assess the safety of input prompts and prevent harmful content generation. A prevailing yet under-examined belief in existing studies is that allocating more computational resources to the safety reasoning budget, i.e., rolling out longer safety thinking trajectories, necessarily yields safer behavior. However, this premise has not been systematically analyzed, despite the widespread adoption of reasoning models.

In this paper, we aim to fill this gap by conducting a rigorous safety evaluation to examine the relationship between the safety reasoning budget and the safety of reasoning model generations, particularly under adversarial conditions. Our analysis reveals that safety does not monotonically improve with longer thinking trajectories: both over-thinking on simple attacks and under-thinking on difficult attacks incur excess safety risks, leading to systematic failures under fixed-budget safety policies. Motivated by these key findings, we propose SafeAdapt: Safety Alignment with adaptive thinking Allocation, a novel approach that learns to dynamically adjust reasoning models' safety thinking budget based on prompt difficulty. This enables reasoning models to allocate computational resources adaptively, rather than following a single fixed thinking pattern (e.g., the longer the safer), thereby mitigating excess risks incurred by under-/over-thinking. Experiments across multiple adversarial benchmarks demonstrate that SafeAdapt significantly improves safety performance with a more ideal thinking budget under diverse and challenging attack settings.

TopFeaRe: Locating Critical State of Adversarial Resilience for Graphs Regarding Topology-Feature Entanglement

Xinxin Fan, State Key Laboratory of AI Safety, Institute of Computing Technology, Chinese Academy of Sciences; and University of Chinese Academy of Sciences; Wenxiong Chen, Dalian University of Technology; and State Key Laboratory of AI Safety, Institute of Computing Technology, Chinese Academy of Sciences; Quanliang Jing, State Key Laboratory of AI Safety, Institute of Computing Technology, Chinese Academy of Sciences; Chi Lin, Dalian University of Technology; Shaoye Luo, State Key Laboratory of AI Safety, Institute of Computing Technology, Chinese Academy of Sciences; and University of Chinese Academy of Sciences; Wenbo Song, Dalian University of Technology; and State Key Laboratory of AI Safety, Institute of Computing Technology, Chinese Academy of Sciences; Yunfeng Lu, Beihang University

Available Media

Graph adversarial attacks are usually produced from the two perspectives of topology/structure and node feature, both of them represent the paramount characteristics learned by today's deep learning models. Although some defense countermeasures are proposed at present, they fails to disclose the intrinsic reasons why these two aspects necessitate and how they are adequately fused to co-learn the graph representation. Towards this question, we in this paper propose an adversarial defense approach through locating the graph's critical state of adversarial resilience, resorting to the equilibrium-point theory in the discipline of complex dynamic system (CDS). In brief, our work has three novelties: i) Adversarial-Attack Modeling, i.e. map a graph regime into CDS, and use the oscillation of dynamic system to model the behavior of adversarial perturbation; ii) 2D Topology-Feature-Entangled Function Design for Perturbed Graph, i.e. project graph topology and node feature as two characteristic spaces, and define two-dimensional entangled perturbation functions to represent the dynamic variance under adversarial attacks; and iii) Location of Critical State of Adversarial Resilience, i.e. utilize the equilibrium-point theory to locate the graph's critical state of attack resilience resorting to the perturbation-reflected 2D function. Finally, multi-facet experiments on five commonly-used realistic datasets validate the effectiveness of our proposed approach, and the results show our approach can significantly outperform the state-of-the-art baselines under four representative graph adversarial attacks.

GoodVibe: Security-by-Vibe for LLM-Based Code Generation

Maximilian Thang, Technical University of Darmstadt; Lichao Wu, University of Bristol; Sasha Behrouzi and Mohamadreza Rostami, Technical University of Darmstadt; Jona te Lintelo, Radboud University; Stjepan Picek, University of Zagreb and Radboud University; Ahmad-Reza Sadeghi, Technical University of Darmstadt

Available Media

Large language models (LLMs) are increasingly used for code generation in fast, informal development workflows, often referred to as vibe coding, where speed and convenience are prioritized, and security requirements are rarely made explicit. In this setting, models frequently produce functionally correct but insecure code, creating a growing security risk. Existing approaches to improving code security rely on full-parameter fine-tuning or parameter-efficient adaptations, which are either costly and prone to catastrophic forgetting or operate at coarse granularity with limited interpretability and control. We present GoodVibe, a neuron-level framework for improving the security of code language models by default. GoodVibe is based on the key insight that security-relevant reasoning is localized to a small subset of neurons. We identify these neurons using gradient-based attribution from a supervised security task and perform neuron-selective fine-tuning that updates only this security-critical subspace. To further reduce training cost, we introduce activation-driven neuron clustering, enabling structured updates with minimal overhead. We evaluate GoodVibe on six LLMs across security-critical programming languages, including C++, Java, Swift, and Go. GoodVibe substantially improves the security of generated code while preserving general model utility, achieving up to a 2.5× improvement over base models, achieving performance competitive with full fine-tuning while using over 4700× fewer trainable parameters, and reducing training computation by more than 3.6× compared to the parameter-efficient baseline (LoRA). Our results demonstrate that neuron-level optimization offers an effective and scalable approach to securing code generation without sacrificing generality.

A Distortion-minimization Watermarking Framework for Large Language Models: Larger Capacity, Stronger Robustness and Higher Quality

Liming Zhai, Xuezhou Shang, Liyun Zhang, and Po Hu, Central China Normal University

Available Media

Large language model (LLM) watermarking provides verifiable source identification for generated text, and its practical deployment requires large watermark capacity, strong robustness against attacks, and high text quality. However, existing methods often struggle to balance all these criteria, typically addressing them with separate designs. To overcome this, we propose a distortion-minimization watermarking (DMW) framework that unifies capacity, robustness and quality within a single optimization paradigm. This framework models robustness and quality as distortion costs for text modifications, minimizing the total distortion for a given watermark length to achieve an optimal trade-off. Specifically, we design several distortion costs: a robustness cost leveraging semantic invariance to resist attacks, and two quality costs guiding modifications toward low-cohesion, high-variability regions to reduce perceptual impact. We then propose periodically optimized syndrome-trellis codes (PO-STCs), formulating overall distortion minimization as a periodic shortest-path problem. This enables real-time optimization for sequential generation with flexible capacity control. Extensive experiments across diverse datasets and LLMs demonstrate DMW's superiority, outperforming state-of-the-art methods across all criteria. Notably, under severe paraphrasing attacks, DMW achieves a match rate up to 46.35% higher than the best baseline, while maintaining superior text quality.

Adversarial Patch EXterminator: Zero-Shot and Patch-Agnostic Defense Framework Against Adversarial Patch Attacks

Jiayimei Wang, City University of Hong Kong; Tao Ni, King Abdullah University of Science and Technology; Guowen Xu, University of Electronic Science and Technology of China; Qingchuan Zhao and Cong Wang, City University of Hong Kong

Available Media

Adversarial patch attacks pose a serious threat to modern computer vision systems. Although existing defense solutions attempt to mitigate such attacks by developing certifiable models or patch identification pipelines, they generally rely on prior knowledge or extensive training data, show insufficient robustness across varying physical conditions, and present limited performance against challenging cases (e.g., tiny, irregular, or highly background-coherent patches). To address such limitations, we propose APEX, a zero-shot, patch-agnostic three-stage adversarial patch defense framework. Specifically, APEX first concentrates patch regions through bounding-box extraction, then integrates a mutual information-based blur heatmap with an edge-aware boundary heatmap to locate adversarial regions, and finally leverages structure-guided image inpainting to restore the image. Our experiments on multiple datasets and existing state-of-the-art defense methods demonstrate that APEX can effectively defend against various types of adversarial patches (e.g., non-naturalistic, naturalistic, and infrared images). In addition, APEX shows superior capability in patch localization, maintains high robustness against varying environments (e.g., lighting conditions) and extreme cases, and also demonstrates high performance in protecting various models in physical-world scenarios.

Attesting Model Lineage by Consisted Knowledge Evolution with Fine-Tuning Trajectory

Zhuoyi Shang, Jiasen Li, and Pengzhen Chen, Institute of Information Engineering, Chinese Academy of Sciences; School of Cyber Security, University of Chinese Academy of Sciences; and State Key Laboratory of Cyberspace Security Defense; Yanwei Liu, Institute of Information Engineering, Chinese Academy of Sciences; and State Key Laboratory of Cyberspace Security Defense; Xiaoyan Gu, Institute of Information Engineering, Chinese Academy of Sciences; School of Cyber Security, University of Chinese Academy of Sciences; and State Key Laboratory of Cyberspace Security Defense; Weiping Wang, Institute of Information Engineering, Chinese Academy of Sciences

Available Media

The fine-tuning technique in deep learning gives rise to an emerging lineage relationship among models. This lineage provides a promising perspective for addressing security concerns such as unauthorized model redistribution and false claim of model provenance, which are particularly pressing in open-weight model libraries where robust lineage verification mechanisms are often lacking. Existing approaches to model lineage detection primarily rely on static architectural similarities, which are insufficient to capture the dynamic evolution of knowledge that underlies true lineage relationships. Drawing inspiration from the genetic mechanism of human evolution, we tackle the problem of model lineage attestation by verifying the joint trajectory of knowledge evolution and parameter modification. To this end, we propose a novel model lineage attestation framework. In our framework, model editing is first leveraged to quantify parameter-level changes introduced by fine-tuning. Subsequently, we introduce a novel knowledge vectorization mechanism that refines the evolved knowledge within the edited models into compact representations by the assistance of probe samples. The probing strategies are adapted to different types of model families. These embeddings serve as the foundation for verifying the arithmetic consistency of knowledge relationships across models, thereby enabling robust attestation of model lineage. Extensive experimental evaluations demonstrate the effectiveness and resilience of our approach in a variety of adversarial scenarios in the real world. Our method consistently achieves reliable lineage verification across a broad spectrum of model types, including classifiers, diffusion models, and large language models.

Identifying Provenance of Generative Text-to-Image Models

Anna Yoo Jeong Ha, Wenxin Ding, Stanley Wu, Shawn Shan, Haitao Zheng, and Ben Y. Zhao, University of Chicago

Available Media

Fine-tuning provides a fast and cheap way to produce new text-to-image models that are often indistinguishable from ones trained from scratch. Unfortunately, misrepresentation of fine-tuned models creates problems for AI companies and users alike, by disincentivizing competition and misleading users on model quality and ethics of its training process.

In this paper, we propose a model provenance system that identifies models produced by fine-tuning on existing text-to-image models, using only black-box query access. Our design is informed by analysis showing that one can quantify the feature space difference between text-to-image models by analyzing their responses to detailed prompts. Our system analyzes model output, extracts visual features using a generic feature extractor, and compares their distributions against those from a reference pool of base models using Jensen-Shannon divergence. Applying statistical hypothesis testing then determines if a target model is trained from scratch or fine-tuned, and if the latter, the likely base (parent) model. We evaluate our system across seven widely used diffusion models and numerous fine-tuned variants. Our results show high accuracy in attributing model lineage, even under adversarial conditions such as image post-processing or weight perturbations. Finally, we demonstrate real world efficacy of our system by tracing provenance of in-the-wild models from popular online platforms.

Track 5

Enigma Track: Attack Detection

Session Chair: David Freeman, Delphi Research

Dover Room

Red-Teaming for Scraping

Malya Jain

Available Media

Scraping has evolved from a nuisance into a sophisticated, economically motivated threat to user privacy and platform integrity. Most organizations rely on reactive rate-limiting and passive monitoring defenses routinely bypassed by well-resourced attackers. This talk presents a different approach: proactively attacking your own platform through a dedicated Scraping Red Team program. We introduce adversary economics as the primary defensive metric such as, the attacker's cost-per-record. When that number is low, your defenses are failing regardless of what your dashboards say. Drawing on real operational experience, we walk through engagement design, privacy-aware data handling, structured vulnerability reporting, and validated remediation.

Malya Jain is a Senior Security Engineer at Meta, where she leads the Vector Discovery team within the Anti-Scraping organization, protecting billions of users. She manages three red teams conducting continuous offensive scraping operations and has built programs spanning static analysis, scraping threat intelligence, and bug bounty integration for scraping vectors. Malya authored "Best Practices for Unauthorized Scraping Red Teaming Exercises," published through the Mitigating Unauthorized Scraping Alliance (MUSA). She has presented at the Underground Economy Conference (France), Ekoparty (LATAM), and RISE USA. Malya is an IEEE Senior Member, WiCyS National Mentor, and holds three GIAC certifications and a Stanford Advanced Cybersecurity Certificate.

Security Tools Don't Detect Attacks—Humans Do

Krupa Brahmkstri, Visa

Available Media

Modern security tools claim to detect attacks, yet in practice they generate signals: alerts, anomalies, and observations. The actual act of detection is performed by human analysts who correlate fragmented evidence, form hypotheses, and reconstruct attack narratives across identities, systems, and time. This distinction exposes a fundamental architectural flaw in modern security operations. Most security infrastructures are designed around alert generation, while the critical work of investigative reasoning remains largely manual.

This talk introduces the concept of Inference-Driven Detection, a model that reframes attack detection as a reasoning problem rather than an anomaly-identification problem. Drawing on lessons from large-scale enterprise security environments, it explores why attacks such as credential abuse, privilege escalation, and lateral movement are often discovered through narrative reconstruction rather than isolated alerts. The talk argues that future security systems must move beyond alert-centric architectures and instead support human and AI-assisted investigative reasoning to identify and understand adversarial behavior.

Krupa Pratap Brahmkstri is a Senior Data Science Manager at Visa, where she leads the development of AI-driven cybersecurity solutions focused on identity risk modeling, insider threat detection, and security operations automation. She builds production-scale machine learning systems that analyze large volumes of enterprise security telemetry to detect anomalous behavior. She brings expertise across artificial intelligence, cybersecurity analytics, and large-scale data systems, with a focus on identity-centric defense strategies and explainable AI for security operations. As a professional mentor and speaker, she is deeply motivated by the opportunity to make a tangible impact through technology. She approaches challenges as mathematical and algorithmic problems, combining analytical rigor with creative problem-solving.

Supply Chain Attacks on Open Source: What's Happening, What Can We Do Today, and What's Next

Zach Steindler, GitHub

Available Media

There's a new pattern of supply chain attacks on open source, where malware exfiltrates secrets from build processes in order to publish more malicious packages from attacker-controlled machines to propagate the attack. These attacks happen at the intersection of how open source projects are maintained, what security capabilities build platforms and package repositories have, what steps consumers of open source take to protect themselves, and how ready we all are to do incident response when an attack happens.

But there are things we can do today to make things better. We'll cover how to secure your build process including removing secrets altogether, how to prepare for an attack by tracking your open source dependencies and putting together an incident response playbook, and where to go for information during an attack and how to minimize impact. Lastly we'll talk about potential signals to proactive detect when a package might be compromised.

Zach is a Principal Security Engineer at GitHub, where he helps secure open source users and enterprise customers. In the OpenSSF he chairs the Technical Advisory Council and co-chair of the Securing Repositories Working Group which helps coordinate security improvements in programming language package repositories like PyPI and npm. Away from computers he enjoys gardening and welding.

5:30 pm–7:00 pm

Symposium Reception

Harborside Ballroom

Sponsored by Meta

Join us at the USENIX Security '26 Reception, featuring a buffet of hors d'oeuvres, crudité, cheeses, and desserts; drinks; and the chance to connect with speakers, symposium organizers, and your fellow attendees.

Friday, August 14

8:00 am–9:00 am

Continental Breakfast

Grand Foyer North and West

9:00 am–10:30 am

Track 1

Applications of Cryptography 5

Session Chair: William Blair, SpaceX

Grand Ballroom VII-X

SPIRIT: Batch Hintless Single-Server PIR via Stateful Ciphertext Conversion

Zhou Zhang, Ran Mao, Zian Zhao, Haowen Pan, Yunhao Fu, Yizhong Liu, and Dawei Li, School of Cyber Science and Technology, Beihang University; Yi Chen, Bo Zhang, Changrui Ren, and Jin Dong, Beijing Academy of Blockchain and Edge Computing; Zhenyu Guan and Song Bian, School of Cyber Science and Technology, Beihang University

Available Media

This work proposes a single-server private information retrieval (PIR) protocol for the efficient evaluation of batched queries without offline hints. It has recently been shown that hintless PIR can be established through the cross-scheme ciphertext conversion technique. However, we point out that such a conversion introduces a considerable amount of computational and communication overheads in the batch PIR setting. To tackle the performance challenge, we propose SPIRIT, a batch, hintless, and single-server PIR protocol that leverages stateful ciphertext conversion to simultaneously achieve high-throughput query execution and low communication cost. To enhance computational efficiency, we develop a novel stateful key expansion procedure that amortizes the conversion cost across multiple queries, significantly accelerating the query evaluation speed. To reduce the communication overhead, we derive a new response compression strategy based on a shared intermediate representation, such that a batch of PIR responses can be packed into a single ciphertext efficiently. In experiments, we show that SPIRIT can achieve up to by 3.1×−5.1× speedup in query execution and reduce the communication bandwidth by 2.3×−3.6× compared to state-of-the-art batch PIR protocols.

Your Keywords Know Each Other: Breaking SSE with <1% Leaked Documents

Mingyu Bian, Jiabei Wang, Dandan Xu, Guangyu Huang, and Yongbin Zhou, School of Cyber Science and Engineering, Nanjing University of Science and Technology

Available Media

Searchable symmetric encryption (SSE) schemes inevitably leak search, access, and volume patterns, which adversaries can exploit along with partial knowledge of the data to recover documents and queries. Prior passive attacks rely on document-keyword occurrence matching, but under realistic low leakage, many keywords induce identical patterns, forming large ambiguous groups that sharply limit recovery.

We observe this failure arises from missing relational signals among keywords. Keywords, however, exhibit stable semantic regularities across corpora even when direct co-occurrence is sparse, allowing external linguistic knowledge to augment observed leakage. Based on this insight, we present Whisper, a staged co-occurrence augmented query recovery framework that weaves LLM-synthesized corpora with pre-trained embeddings to densify co-occurrence matrices and disambiguate otherwise indistinguishable candidates.

We evaluate Whisper on six diverse real-world datasets, including three collected in late 2025 to test generalization beyond LLM training cutoffs. At 0.1% leakage, previously considered safe, Whisper boosts correct query recovery from 1727% under current SOTA to 36–49% across all datasets, a 22.5× improvement. Even at 1% leakage, Whisper approaches near-complete recovery, exceeding 80% on most datasets and peaking at 95%, while remaining effective against padding and volume hiding defenses, exposing a previously underestimated vulnerability in SSE.

SONIC: Concurrent Oblivious RAM & Data Structures for Low-Latency and High-Throughput

Nihal Talur and Ioannis Demertzis, UC Santa Cruz

Available Media

Relying solely on encryption for privacy-preserving computations is prone to leakage-abuse/access-pattern attacks. TEEs, while cost-effective, are also vulnerable to side-channel attacks. Oblivious primitives, such as oblivious memory (ORAM) and data structures (ODS) are effective building blocks to mitigate these risks by concealing memory access patterns and side-channel information. Applications range from private contact discovery (Signal) to anonymous key transparency, encrypted email search, encrypted/oblivious databases, anonymous communication (Sparta/SP'25), private federated learning, LLM privacy (Compass/OSDI'25), and broader confidential computing efforts.

Tree-based ORAMs (EnigMap (USENIX'23), GraphOS (PVLDB'23), Oblix (SP'18)) offer low latency but limited parallelism, struggling to exceed 1K req/s throughput even for small datasets. Partition-based solutions like Snoopy (SOSP'21) split data into multiple subORAMs, each parallel-scanning its shard via an oblivious hash table built from incoming requests, achieving high throughput by generously trading off latency—theoretically enabling linear scalability. In practice, Snoopy's performance hinges on how quickly each subORAM can complete its sequential scan before exceeding latency targets—constraining server utilization and throughput. While TB-scale datasets are theoretically feasible by adding servers, it requires 1000+ servers in practice.

In this work, we reconcile the aforementioned fractured landscape between low-latency and high-throughput ORAM solutions. We introduce SONIC: the first ORAM for hardware enclaves that replaces PathORAM (used by EnigMap, GraphOS, and Oblix) with RingORAM. Our design is the first low-latency ORAM achieving minimum throughput of 150K req/s and up to 2M req/s (with one server), tackling core challenges of all tree-ORAM constructions (including RingORAM) such as overcoming the sequential eviction bottleneck, enabling efficient batch evictions, and providing lock-free access, reshuffle, and stash operations. SONIC achieves 28-197× higher ORAM access throughput than the open-source EnigMap implementation, and 158-1065× higher than GraphOS (for N=227 and block size 64 bytes). SONIC offers various methods to leverage ORAM concurrency for building oblivious data structures, such as OMAPs. Finally, SONIC can serve as a drop-in replacement for Snoopy's subORAM to provide more practical scalability—1TB can now be handled with just 32 servers instead of thousands.

DDR-SSE: Duplicated Retrieval of Documents for System-wide Secure Searchable Symmetric Encryption

Zichen Gui, University of Georgia; Simon-Philipp Merz and Kenneth G. Paterson, ETH Zürich; Sikhar Patranabis, IBM Research India

Available Media

Searchable Symmetric Encryption (SSE) schemes enable efficient keyword searches over encrypted documents at the cost of some leakage. An SSE scheme is said to be system-wide secure if it resists cryptanalysis by an adversary with access to leakage from retrieval of both encrypted indices and encrypted documents. The vast majority of state-of-the-art SSE schemes are, in fact, not system-wide secure (Gui et al., IEEE S&P 2023). Currently, the only efficient system-wide secure SSE scheme is SWiSSSE (Gui et al., PoPETS 2024). However, SWiSSSE requires a client state that is updated per query (which hinders adoption in various practical settings), and its leakage is hard to characterize precisely (thus making security analysis harder).

In this paper, we present DDR-SSE – a practically efficient, system-wide secure SSE scheme that only requires a static client state, and has a simple leakage profile. Technically, we introduce a novel encrypted document retrieval scheme that uses duplicated document storage and randomized document retrieval to suppress access pattern leakage without compromising on practical efficiency. A remarkable feature of our scheme is its conceptual simplicity (unlike SWiSSSE, which uses an extremely involved document retrieval mechanism).

We present a simulation-based security proof for DDR-SSE with respect to a rigorously formal system-wide leakage profile. Through extensive leakage cryptanalysis, we establish that DDR-SSE is resilient to query reconstruction attacks (even under "unrealistically" strong attack assumptions). Finally, we benchmark a prototype implementation of DDR-SSE and show that it scales smoothly to large databases of the size seen in real-world applications.

mmCipher: Batching Post-Quantum Public Key Encryption Made Bandwidth-Optimal

Hongxiao Wang, The University of Hong Kong; Ron Steinfeld, Monash University; Markku-Juhani O. Saarinen, Information Security Laboratory, Tampere University; Muhammed F. Esgin, Monash University; Siu-Ming Yiu, The University of Hong Kong

Available Media

In applications such as secure group communication and broadcasting, it is important to efficiently deliver multiple messages to different recipients at once. To this end, multi-message multi-recipient Public Key Encryption (mmPKE) enables the batch encryption of multiple messages for multiple independent recipients in one go, significantly reducing costs–particularly bandwidth–compared to the trivial solution of encrypting each message individually. This capability is especially desirable in the post-quantum setting, where the ciphertext length is typically significantly larger than the corresponding plaintext. However, almost all prior works on mmPKE are limited to quantum-vulnerable traditional assumptions.

In this work, we propose the first CPA-secure mmPKE and Multi-Key Encapsulation Mechanism (mmKEM) from the standard Module Learning with Errors (MLWE) lattice assumption, named mmCipher-PKE and mmCipher-KEM, respectively. Our design proceeds in two steps: (i) We introduce a novel generic construction of mmPKE by proposing a new PKE variant—extended reproducible PKE (XR-PKE)—that enables the reproduction of ciphertexts through additional hints; (ii) We instantiate a lattice-based XR-PKE using a new technique that can precisely estimate the impact of such hints on the ciphertext security while also establishing suitable parameters. We believe both to be of independent interest. As a bonus contribution, we explore generic constructions of adaptively secure mmPKE, resisting adaptive corruption and chosen-ciphertext attacks.

We also provide an efficient implementation and thorough evaluation of the practical performance of our mmCipher. The results demonstrate substantial bandwidth and computational savings over the state-of-the-art. For example, for 1024 recipients, our mmCipher-KEM achieves a 23–45× reduction in bandwidth overhead, with ciphertexts only 4–9% larger than the plaintexts (near optimal bandwidth), while also offering a 3–5× reduction in computational cost.

ZipPIR: High-throughput Single-server PIR without Client-side Storage

Rasoul Akhavan Mahdavi, Abdulrahman Diaa, and Florian Kerschbaum, University of Waterloo

Available Media

Private Information Retrieval (PIR) allows a client to privately access a database without revealing which element is accessed. Initial PIR protocols based on Ring Learning with Errors (RLWE) demonstrated the practicality of PIR, but achieve limited throughput. Alternatively, high-throughput protocols leverage an offline phase that requires substantial client-side storage (e.g., hints in SimplePIR) or involve prohibitive communication costs during the offline phase (e.g., Piano). These limitations conflict with the practical constraints of resource-limited clients and are further exacerbated by dynamic databases, where updates necessitate costly regeneration and retransmission of hints.

To address these challenges, we propose ZipPIR, a high-throughput PIR protocol that compresses LWE ciphertexts into significantly smaller Paillier ciphertexts. ZipPIR leverages the offline phase to obtain this size reduction without incurring the associated computational cost in the online phase. Moreover, under computational assumptions, ZipPIR features an almost silent offline phase, requiring no communication beyond an initial public key, enabling the server to independently generate and update hints during idle times without client interaction. ZipPIR achieves over 2 GB/s of throughput — comparable to state-of-the-art protocols such as SimplePIR — without the need for a large client-stored hint. For PIR over a 1 GB database, ZipPIR has up to 10x higher throughput than existing protocols with no client-side storage, while requiring less than 200 KB of server-side storage per client, significantly enhancing scalability for practical deployments. While prior PIR protocols using Paillier are very inefficient, ZipPIR is the first PIR protocol using Paillier that achieves throughput that is competitive with state-of-the-art PIR protocols.

Shadowfax: Hybrid Security and Deniability for AKEMs

Phillip Gajland, IBM Research Europe – Zurich; Vincent Hwang, MPI-SP, Radboud University; Jonas Janneck, Ruhr University Bochum

Available Media

As cryptographic protocols transition to post-quantum security, most adopt hybrid solutions combining classical and post-quantum assumptions. This shift often sacrifices efficiency, compactness, or even security. One such property is deniability, which enables users to plausibly deny authorship of potentially incriminating messages. While classical protocols like X3DH key agreement (used in Signal and WhatsApp) provide deniability, post-quantum protocols like PQXDH and Apple's iMessage with PQ3 do not.

This work addresses this gap by investigating how to efficiently preserve deniability in post-quantum protocols. Specifically, we propose two hybrid schemes for authenticated key encapsulation mechanisms (AKEMs). The first is a black-box construction that preserves deniability when both constituent AKEMs are deniable. The second is Shadowfax, a non-black-box AKEM that achieves hybrid security, integrating a classical non-interactive key exchange, a post-quantum key encapsulation mechanism, and a post-quantum ring signature. Shadowfax satisfies deniability in both dishonest and honest receiver settings, relying on statistical security in the former and on a single pre- or post-quantum assumption in the latter.

Finally, we provide several portable implementations of Shadowfax. When instantiated with standardised components (ML-KEM and Falcon), Shadowfax yields ciphertexts of 1728 bytes and public keys of 2036 bytes, with encapsulation and decapsulation costs of 1.8M and 0.7M cycles on an Apple M1 Pro.

Non-Interactive Key Exchange from Lattices in the Standard Model

Ying Li and Lei Zhang, East China Normal University; Qiqi Lai, Shaanxi Normal University

Available Media

Existing non-interactive key exchange (NIKE) constructions in the CKS-heavy security model either incur loose security reductions or lack lattice instantiations due to the absence of suitable hash proof systems (HPS). To overcome this limitation, we introduce a new cryptographic primitive, named as decoupled key encapsulation mechanisms (decoupled KEMs), which separates recipient-independent ciphertext generation from recipient-specific key derivation. Moreover, we formalize the security of our decoupled KEM as MU-IND-wCCA^Corr+, which is a multi-user security notion. Based on this, we give a new framework that any decoupled KEM with MU-IND-wCCA^Corr+ implies a NIKE protocol with much tighter security in the CKS-heavy security model. For concrete instantiation, we first construct a specific HPS based on Modulus-LWE in the standard model, and then obtain a decoupled KEM with the desired security notion. Applying our framework, we obtain a lattice-based NIKE secure in the standard model, denoted NIKEHPS. NIKEHPS. achieves a linear security reduction loss in the number of users and close to the known optimum, improving upon the quadratic loss of SWOOSH (USENIX Security' 24) under CKS-heavy security. Experiments show that our protocol achieves significant speedups in key generation and shared-key derivation over SWOOSH.

DuetORAM: Two-Server Distributed ORAM with Constant Rounds and O(log N) Communication

Feng Li and Xiangfu Song, Nanyang Technological University; Yingying Li, Hangzhou Normal University; Lisha Yao and Guomin Yang, Singapore Management University; Tianwei Zhang, Nanyang Technological University; Robert H. Deng, Singapore Management University

Available Media

Distributed Oblivious RAM (DORAM) is a promising building block for privacy-preserving cloud databases and outsourced storage systems. However, existing two-server designs often rely on slow linear scans or heavy cryptographic primitives, making them struggle to balance efficiency and bandwidth, and thus hindering their practical deployment.

We present DuetORAM, a two-server DORAM that achieves constant-round access with O(log N) communication while avoiding these computational bottlenecks. Our key idea is a replicated-to-shared block encoding that allows servers to keep identical ciphertexts for efficient PIR-based retrieval, while locally interpreting them as secret shares to enable oblivious eviction via a lightweight shuffle. We further design a secret-shared shuffle with an offline-online decomposition that shifts most bandwidth-intensive work to a preprocessing phase, significantly reducing online communication. We implement a prototype of DuetORAM and evaluate it under diverse network conditions. Our results show that DuetORAM outperforms both the state-of-the-art two-server scheme DUORAM (reducing retrieval latency by up to 170× in LAN settings), and three-server design S^3ORAM (reducing retrieval latency by 1.7× in LAN and accelerating eviction by 7× in LAN and 5× in WAN, respectively).

Track 2

Security of ML 5

Session Chair: Omer Akgul, RSAC Labs

Grand Ballroom VI

HijackKV: New Threat in Position-Independent KV Cache Reuse

Yichi Zhang and Zhiqi Wang, The Pennsylvania State University; Huan Zhang, University of Illinois Urbana-Champaign; Yuchen Yang, The Pennsylvania State University

Available Media

Key-Value (KV) cache reduces inference latency in large language models (LLMs). Traditional prefix-based reuse has low cache hit rates across inference requests because it requires exact token and position matches. To improve efficiency, recent system optimizations introduce position-independent KV reuse, allowing KV cache to be reused whenever identical text chunks appear, regardless of their position in the sequence.

We show this design introduces a new threat, KV Cache Hijacking. Since KV caches are retrieved by token match but encode the context in which they were originally computed, the KV tied to a benign-looking token chunk may encode an attacker-controlled prefix. When later reused in a victim query, this contaminated KV silently hijacks the model's behavior, even if no attacker-controlled text appears in the input.

We introduce HIJACKKV, the first attack framework that systematically exploits this vulnerability, demonstrating its severity and practicality. HIJACKKV optimizes an attacker-controlled prefix, so that the KV computed for a subsequent common benign text encodes the attacker's goal, while the text remains unchanged for future cache hits. HIJACKKV achieves an average 94% success rate in a single attempt, remains effective under realistic constraints including low hit rates (10%) and frequent recomputation (50%), persists over multi-turn interactions, and transfers across models in black-box settings. We further provide design insights for building secure KV reuse systems.

MarkNull: Model-Agnostic Watermark Removal in AI-Generated Images via On-Manifold Latent Manipulation

Jie Cao, Qi Li, Zelin Zhang, and Xiaodong Wu, Queen's University; Lingshuang Liu, University of Waterloo; Xiangman Li and Jianbing Ni, Queen's University

Available Media

Digital watermarking has emerged as a critical technique for provenance and copyright attribution in AI-generated imagery, yet its robustness against realistic, model-agnostic removal attacks remains poorly explored. Existing attacks either succeed only against specific generative models or achieve removal at the cost of severe visual degradation. In this paper, we propose MarkNull, a model-agnostic watermark removal attack via on-manifold latent manipulation. MarkNull is grounded in a key observation: watermarked images exhibit a strong statistical dependency between the generated latent representation and the embedded initial noise. To quantify this dependency, we introduce the Noise-Latent Alignment Score (NLAS) and formulate an optimization objective that selectively decorrelates the latent representation from the embedded watermark while preserving semantic fidelity. Extensive evaluations across different categories of watermarking paradigms, including post-hoc, fine-tuning-based, and initial-noise-based schemes, demonstrate that MarkNull reduces average bit accuracy to 53.14%, approaching random-guessing (50%), without perceptible image degradation. To further improve scalability, we propose MarkNull-A, an amortized, optimization-free variant that distills the attack into a single forward pass, achieving 0.50 s/image with modest computational overhead. Notably, our attacks successfully compromise Google's SynthID-Image system while preserving high visual quality and transfer effectively to video watermarking. Finally, we present an attack detection mechanism as a defensive counterpart to MarkNull and MarkNull-A, highlighting the necessity of developing watermark designs resilient to model-agnostic latent-space attacks.

BadGraph: Structural Knowledge Isolation Attacks against Graph Retrieval-Augmented Generation

Leiming Yan, Xinlong Xu, and Ziqiang Li, Engineering Research Center of Digital Forensics, Ministry of Education, Nanjing University of Information Science & Technology

Available Media

Graph Retrieval-Augmented Generation (GraphRAG) improves cross-document reasoning by introducing graph structures into retrieval. However, these graph structures also expose an under-studied topological attack surface. Existing research on GraphRAG attacks primarily focuses on targeted attacks based on text or relation manipulation, which induce the model to generate specific incorrect answers by injecting factual errors. These attacks often rely on explicit false-answer payloads, poisoned relations, or corpus rewriting, and their documents can be easier to flag during content auditing.

In this paper, we propose a targeted availability attack against GraphRAG: Structural Knowledge Isolation. Unlike traditional wrong-answer manipulation attacks, this attack targets the system's availability by degrading retrieval of critical evidence through topological manipulation. The resulting context can lack the evidence needed for grounded answer generation. We first mathematically analyze three topological vulnerabilities of graph algorithms used in GraphRAG systems when subjected to topological perturbations. Based on this analysis, we propose the BadGraph attack framework. By injecting a small amount of semantically neutral text, it generates adversarial subgraphs in the knowledge graph and reduces the visibility of target evidence in top-ranked retrieval contexts. Experiments show that with a small document injection budget (49 documents on HotpotQA and 30 on 2WikiMultiHopQA), BadGraph consistently lowers source recall and end-to-end QA F1 on three mainstream GraphRAG systems (MS-GraphRAG, LightRAG, and FastGraphRAG), while using neutral linker documents.

FragFuse: Bypassing Access Control of Large Language Model Agents via Memory-Based Query Fragmentation and Fusion

Zixin Rao, Wentian Zhu, and Chan Aristella Lu, University of Georgia; Zhaorun Chen, University of Chicago; Wei Niu and Le Guan, University of Georgia; Bo Li, University of Illinois Urbana-Champaign; Zhen Xiang, University of Georgia

Available Media

Large language model (LLM) agents increasingly rely on long-term memory to support complex task execution, user personalization, and domain adaptation. Meanwhile, emerging access-control mechanisms for LLM agents are being explored to block policy-violating requests, aiming to prevent misuse and improve resource efficiency. In this paper, we reveal a novel attack surface arising from agents' memory operations: prohibited content triggering access control can be fragmented across interactions, stored in long-term memory in a benign-appearing form, and later reconstructed through memory retrieval, without appearing explicitly in the final user query. Specifically, we propose FragFuse, the first attack that enables unprivileged users to bypass agent access control by exploiting this temporal channel introduced by long-term memory. FragFuse operates in three stages: (1) identifying rejection-responsible fragments via black-box adaptive querying with fragment masking; (2) injecting these fragments into memory using marked carrier queries; and (3) retrieving and fusing the stored fragments through a follow-up attack query. While FragFuse can be instantiated manually for individual agents, we propose an optimization scheme that tunes fusion instructions and marker designs on surrogate models, enabling automated attack generation without violating the attacker's threat model assumptions. We evaluate FragFuse across four representative agent settings and task domains, covering three state-of-the-art agent access-control mechanisms. FragFuse achieves an average bypass success rate of 86.3% and an average end-to-end harmful task success rate of 41.1% across all settings, with only 4.4% average task success rate degradation compared to configurations without access control. Additionally, we show that alternative defenses, such as state-of-the-art prompt-injection detectors and perplexity detectors, cannot effectively address our attack.

From Zero to Hero: Cross-modal-enhanced Adversarial Item Promotion Attack against Multimodal Recommender Systems

Mengyu Yao, MOEKey Lab of HCST (PKU), School of Computer Science, Peking University; Ant Group; Ziqi Zhang, University of Illinois Urbana-Champaign; Yifeng Cai and Junlin Liu, MOEKey Lab of HCST (PKU), School of Computer Science, Peking University; Xinyi Fu and Weiqiang Wang, Ant Group; Xiangqun Chen, Yao Guo, and Ding Li, MOEKey Lab of HCST (PKU), School of Computer Science, Peking University

Available Media

Multimodal recommender systems (MRSs) jointly leverage visual and textual item representations to determine item ranking and exposure in modern online platforms. Their strong dependence on item content, however, introduces new security risks. In particular, malicious sellers can conduct the adversarial item promotion (AIP) attack to manipulate item content to deceive the recommender into overranking specific items. When it succeeds, the promoted items gain disproportionately higher exposure, leading to significant market visibility and direct economic gain. However, existing AIP research primarily targets unimodal recommenders, leaving the unique vulnerabilities of MRSs largely unexplored. Meanwhile, multimodal adversarial attacks for vision–language models (VLMs) optimize objectives unrelated to ranking and lack cross-modal coordination unique to MRSs. To bridge this gap, we propose CREAM (CRoss-modal-Enhanced AIP attack against MRSs). Our key insight is to jointly perturb multiple modalities in a semantically consistent manner}. We integrate a tailored visual perturbator, a text generator, and a joint optimization controller to fully exploit cross-modal correlations in a black-box setting. Our comprehensive evaluation shows that CREAM significantly outperforms existing methods, achieving on average 5.75x and 2.89x higher exposure at top-10 and top-50 metrics, and demonstrates robustness under evolving real-world conditions. This exposes a tangible economic risk to recommender platforms. Meanwhile, CREAM maintains high imperceptibility across visual, textual, and cross-modal dimensions. We further investigate several potential defense strategies and demonstrate their limitations, highlighting the urgent need for stronger protections against adversarial threats in MRSs.

On Evaluating the Robustness of Large Vision-Language Models via Untargeted Modality Alignment Breaking Adversarial Attack

Zhichao Li, Hongshan Yang, Zhibo Wang, Huiyu Xu, and Junhong Lai, Zhejiang University; Yaopeng Wang, Southeastern University and Zhejiang University; Kui Ren and Chun Chen, Zhejiang University

Available Media

Large Vision-Language Models (LVLMs) have achieved remarkable success in multimodal tasks by aligning the representation space of visual encoders to that of the LLMs. However, they remain vulnerable to transferable adversarial attacks, which can manipulate the LVLMs' output without accessing the model. Ensuring their reliable deployment thus requires a rigorous evaluation of black-box robustness. Current methods provide a limited assessment by perturbing only the visual encoder of LVLMs and often neglect untargeted attack scenarios. In this work, we propose the Modality Alignment Breaking Attack (MABA), a novel transferable, untargeted adversarial attack for evaluating the black-box robustness of LVLMs. MABA emphasizes disrupting the entire multimodal pipeline, targeting two key phases: visual encoding and modality alignment. First, MABA reveals that the core of transferable adversarial attacks lies in suppressing discriminative visual representations and explicitly uses this as an optimization objective to improve transferability across different LVLMs. Second, MABA introduces a mutual-information-aware projector that acts as a surrogate modality alignment module of LVLMs, effectively breaking cross-modal consistency and enhancing the transferability. Extensive evaluations demonstrate that MABA achieves state-of-the-art performance, leading to an average 58.37% drop in semantic metrics for the image caption task. Through ablation studies on diverse LVLM families, we derive valuable insights into strengthening the robustness of LVLMs.

Attacks on Approximate Caches in Text-to-Image Diffusion Models

Desen Sun, Shuncheng Jie, and Sihang Liu, University of Waterloo

Available Media

Diffusion models are a powerful class of generative models that produce images and other content from user prompts, but they are computationally intensive. To mitigate this cost, recent academic and industry work has adopted approximate caching, which reuses intermediate states from similar prompts in a cache. While efficient, this optimization introduces new security risks by breaking isolation among users. This paper provides a comprehensive assessment of the security vulnerabilities introduced by approximate caching. First, we demonstrate a remote covert channel established with the approximate cache, where a sender injects prompts with special keywords into the cache system and a receiver can recover that even after days, to exchange information. Second, we introduce a prompt stealing attack using the approximate cache, where an attacker can recover existing cached prompts from hits. Finally, we introduce a poisoning attack that embeds the attacker's logos into the previously stolen prompt, leading to unexpected logo rendering for the requests that hit the poisoned cache prompts. These attacks are all performed remotely through the serving system, demonstrating severe security vulnerabilities in approximate caching. The code for this work is available.

When AIOps Become "AI Oops": Subverting LLM-driven IT Operations via Telemetry Manipulation

Dario Pasquini, RSAC Labs; Evgenios M. Kornaropoulos and Giuseppe Ateniese, George Mason University; Omer Akgul, Athanasios Theocharis, and Petros Efstathopoulos, RSAC Labs

Available Media

AI for IT Operations (AIOps) is transforming how organizations manage complex software systems by automating anomaly detection, incident diagnosis, and remediation. Modern AIOps solutions increasingly rely on autonomous LLM-based agents to interpret telemetry data and take corrective actions with minimal human intervention, promising faster response times and operational cost savings.

In this work, we perform the first security analysis of AIOps solutions, showing that, once again, AI-driven automation comes with a profound security cost. We demonstrate that adversaries can manipulate system telemetry to mislead AIOps agents into taking actions that compromise the integrity of the infrastructure they manage. We introduce techniques to reliably inject telemetry data using error-inducing requests that influence agent behavior through a form of adversarial reward-hacking; plausible but incorrect system error interpretations that steer the agent's decision-making. Our attack methodology, AIOpsDoom, is fully automated–combining reconnaissance, fuzzing, and LLM-driven adversarial input generation–and operates without any prior knowledge of the target system.

To counter this threat, we propose AIOpsShield, a defense mechanism that sanitizes telemetry data by exploiting its structured nature and the minimal role of user-generated content. Our experiments show that AIOpsShield reliably blocks telemetry-based attacks without affecting normal agent performance. Ultimately, this work exposes AIOps as an emerging attack vector for system compromise and underscores the urgent need for security-aware AIOps design.

Low-Cost Hard-Label Adversarial Attack with Theoretical Foundations

Jun Liu, University of Macau; National Institute of Informatics; Leo Yu Zhang, Griffith University; Fengpeng Li, University of Macau; Isao Echizen, University of Tokyo; National Institute of Informatics; Jiantao Zhou, University of Macau

Available Media

Hard-label black-box attacks, relying solely on top-1 predictions, represent one of the most challenging yet practically threat models. Despite recent progress, existing approaches face two key limitations: (1) they overlook the critical role of initialization, focusing primarily on optimization strategies; and (2) they rely heavily on empirical heuristics without theoretical guarantees. To bridge this gap, we establish a unified theoretical framework showing that existing sign-flipping hard-label attacks can be understood as approximating the true gradient sign. Guided by this {principled analysis}, we propose a novel attack framework featuring a zero-query initialization strategy and a Pattern-Driven Optimization (PDO) algorithm. We provide theoretical guarantees that our initialization yields higher cosine similarity to the true gradient sign than random baselines, and our PDO module achieves significantly lower query complexity than baseline search methods. Extensive experiments across CIFAR-10, ImageNet, and ObjectNet—covering standard and adversarially trained models, commercial APIs, and CLIP models—demonstrate that our method consistently outperforms SOTA hard-label attacks in both success rate and efficiency, particularly under low query budgets. Furthermore, our method demonstrates robust generalization across corrupted data (ImageNet-C), biomedical images (PathMNIST), and dense prediction tasks such as segmentation. Notably, it bypasses the stateful defense Blacklight, achieving a 0% detection rate.

Track 3

Security and Privacy for Web, Mobile, and Emerging Technologies 3

Session Chair: Aurore Fass, Inria Centre at Université Côte d’Azur

Grand Ballroom V

BlockMeNot: Automatic Selection of Domain and URL Blocking Granularity to Minimize Collateral Damage and Evasion

Daud Ahmed, IMDEA Software Institute and Universidad Politécnica de Madrid; Srdjan Matic, Universidad Carlos III de Madrid; Platon Kotzias and Emiliano Carlesi, BforeAI; Juan Caballero, IMDEA Software Institute

Available Media

Domain and URL blocking is a fundamental mechanism for mitigating malicious, illegal, and inappropriate online content. Yet, selecting the right blocking granularity remains a largely unexplored problem. Overblocking can cause severe collateral damage if the content to be blocked uses the infrastructure of benign services, which may be unintentionally disrupted. Underblocking instead enables trivial evasion. This paper addresses this gap by proposing a novel approach that, given a domain or URL to be blocked, automatically selects the most appropriate blocking granularity (URL, subdomain, or apex) to minimize collateral damage and evasion opportunities. At the core of our approach are two novel machine learning classifiers to identify URL-leasing and subdomain-leasing apexes. The classifiers are trained and evaluated using a manually labeled dataset of 10,843 apexes, achieving F1 scores over 0.9. We implement our approach into BlockMeNot, a tool for identifying potential collateral damage prior to enforcement. We evaluate BlockMeNot on 225,355 entries (i.e., domains and URLs) collected from six blocklists and two threat exchanges. Although only 2.6% of apexes belong to leasers, they host 34.7% of the entries, and up to 71.7% in phishing-focused blocklists, underscoring the importance of fine-grained blocking. BlockMeNot identifies 1,976 leasing apexes, 59.3% of which were not in our ground truth, demonstrating the limitations of static lists and the need for our classifiers. Only 37.2% of blocklist entries are listed at the optimal granularity, 1.3% produce collateral damage, and 61.6% enable easy evasion. Overall, our work provides the first automated solution for domain and URL blocking granularity selection and offers practical guidance for block requesters, executors, and blocklist maintainers.

H3Act: Automated Measuring Semantic Conversion Anomalies of HTTP/3-to-HTTP/1.1 Translation in CDNs

Qihang Peng and Siyuan Tian, Institute for Network Sciences and Cyberspace, BNRist, Tsinghua University; Yongxin Qiu, Beihang University; Jinyang Huang, Central South University; Yaru Yang, Institute for Network Sciences and Cyberspace, BNRist, Tsinghua University; Xiang Li, Nankai University; Zhongguancun Laboratory; Jia Zhang, Institute for Network Sciences and Cyberspace, BNRist, Tsinghua University; Zhongguancun Laboratory; Yiming Zhang, Institute for Network Sciences and Cyberspace, BNRist, Tsinghua University; Haixin Duan, Institute for Network Sciences and Cyberspace, BNRist, Tsinghua University; Quancheng Laboratory; Yunsenxiao Lin and Shugen Chen, Tencent Technology Co., Ltd.; Liqun Yang, Beihang University

Available Media

Content Delivery Networks (CDNs) are adopting HTTP/3 to enhance performance; however, they often need to convert it to HTTP/1.1 for compatibility. This conversion creates significant attack surfaces and may reintroduce confirmed or even patched vulnerabilities in HTTP/2 or HTTP/1. Unfortunately, existing tools struggle to adapt to the HTTP/3 environment and efficiently leverage accumulated attack knowledge. To overcome these challenges and systematically measure HTTP/3-to-HTTP/1.1 conversion anomalies within the black-box CDN environment, we present H3Act, a dual-agent, knowledge-driven fuzzing framework targeting HTTP/3-to-HTTP/1.1 semantic conversion anomalies. Our approach combines Large Language Models (LLMs) with Hybrid Retrieval-Augmented Generation (RAG) to automatically transform protocol specifications and historical threat intelligence into high-precision HTTP/3 test payloads, enabling regression testing of semantic translation risks. In a large-scale study of 9 commercial CDNs, including Cloudflare, Cloudfront, and Tencent CDN, we found a significant regression in protocol security, which means vulnerabilities in old protocols are reintroduced in HTTP/3. Our research identified 7 common attack vectors across various categories, including request smuggling, cache poisoning, and Denial-of-Service (DoS) amplification. Every CDN is vulnerable to at least one attack vector. We have responsibly disclosed these vulnerabilities and have received confirmations from some vendors. These findings highlight that the CDN ecosystem currently lacks the capacity to maintain security consistency checks while pursuing improvements in protocol performance.

JScamd: An Automated Static Taint Analysis Framework for Detecting Cryptographic API Misuses in JavaScript

Shijie Jia, State Key Laboratory of Cyberspace Security Defense, Institute of Information Engineering, Chinese Academy of Sciences; School of Cyber Security, University of Chinese Academy of Sciences; Bowen Xu, Faculty of Electrical Engineering and Computer Science, Ningbo University; Yuan Ma, Yingjiao Niu, Limin Liu, and Daren Zha, State Key Laboratory of Cyberspace Security Defense, Institute of Information Engineering, Chinese Academy of Sciences; School of Cyber Security, University of Chinese Academy of Sciences; Jingqiang Lin, School of Cyber Security, University of Science and Technology of China

Available Media

Cryptographic APIs are essential to application security, yet frequent developer misuse leads to widespread vulnerabilities. Detecting cryptographic misuse in JavaScript is challenging due to its dynamic semantics and fragmented cryptographic library ecosystem, which hinder static analysis tools from accurately modeling data flows and covering non-standard or evolving libraries, leading to high false-negative rates.

In this paper, we propose JScamd, a static analysis framework for scalable detection of cryptographic misuse in JavaScript. JScamd integrates an enhanced program representation engine, an LLM-assisted taint rule generation pipeline, and a precise static taint analysis engine to address JavaScript dynamic features. Building upon 16 systematically defined cryptographic security constraints, JScamd automatically analyzes cryptographic APIs, identifies security-sensitive taint sources and sinks, and synthesizes comprehensive taint analysis rules. This design enables accurate detection of cryptographic misuse across diverse JavaScript libraries and application scenarios. We performed a comprehensive large-scale evaluation on 322 widely adopted JavaScript projects. JScamd detects 1,059 misuses and achieves 97.78% precision, significantly outperforming three state-of-the-art tools. Our analysis further uncovers previously unknown vulnerabilities, resulting in 41 assigned CVEs. These results demonstrate that JScamd effectively bridges cryptographic security specifications and ecosystem-scale vulnerability detection in JavaScript.

When Fun Turns Toxic: A First Look at Aggressive Advertising in Mini-games

Pei Chen, Geng Hong, Yicheng Qin, Huazhe Wang, Mengying Wu, and Min Yang, Fudan University; Ziru Zhao, Yuanpeng Zhu, and Tao Su, vivo Mobile Communication Co., Ltd

Available Media

Mini-games have emerged as a dominant paradigm within super-app ecosystems, enabling lightweight services like casual games to reach millions of users instantly. While official advertisement interfaces simplify monetization, the ease of integration and insufficient oversight have led to aggressive and potentially deceptive advertising practices, severely degrading the user experience. Aggressive advertising, though not malware, still subverts platform security boundaries by abusing legitimate APIs to bypass auditing, manipulate user interaction, and undermine platform trust, constituting a systemic security risk rather than mere policy violation.

In this work, we conduct the first systematic security analysis of aggressive advertising in mini-games. We analyze platform policies and developer capabilities across nine mini-game platforms, and characterize aggressive advertising behaviors. We further design a scalable detection framework, MAAD, and perform a large-scale measurement across three major platforms, i.e., WeChat, Facebook Instant Games, and Quickgame, revealing that 49.95% of mini-games exhibit aggressive advertising, including cases in highly popular titles with over 100k user reviews. Our analysis further uncovers their disruptive behavioral patterns, such as game-specific triggers, excessive pop-up frequency, and misleading strategies, as well as adversarial bypass techniques. These findings uncover that aggressive advertising constitutes a widespread form of platform abuse enabled by structural blind spots in current enforcement mechanisms. We provide actionable implications for strengthening platform governance, detection, and long-term ecosystem resilience.

BUIzz: Finding Policy Enforcement Bugs via Interaction Simulation on the Browser User Interface

Mingi Jung, Donggyu Kim, Mijung Kim, and Seongil Wi, UNIST

Available Media

Modern web ecosystems rely on security policy headers, such as Content Security Policy (CSP) and the SameSite cookie attribute, for client-side defenses. Because browsers enforce these headers, browser bugs in policy enforcement directly undermine these defenses. While recent studies have attempted to find such bugs, they largely overlook bugs triggered by user interactions on the browser user interface (BUI).

In this paper, we propose BUIZZ, the first testing framework that identifies policy enforcement bugs triggered by BUI-level user interactions. BUIZZ first collects a comprehensive set of interactions by referencing browser manuals, right-click context menus, and known browser bugs. It then executes each interaction and their combinations on the test pages via OS-level simulation. Instead of cross-browser differential testing, BUIZZ leverages a pre/post-interaction oracle that checks for enforcement inconsistencies before and after applying interactions, enabling bug detection within a single browser. We demonstrate the efficacy of BUIZZ by finding 35 security bugs and three functional bugs across six browsers, including Chrome and Firefox. Our reports have led to fixes for 14 security bugs, resulting in seven CVEs and $14,700 in bug bounties.

SoK: PHILTER: Uncovering Security and Functional Gaps in AI-based Phishing Website Detection Literature via an LLM-based Reasoning Framework

Mahbub Alam, Texas A&M University; Muhammad Lutfor Rahman, California State University San Marcos; Sonjoy Kumar Paul, Amy W. Hays, Aftab Hussain, Md Imanul Huq, and Nitesh Saxena, Texas A&M University

Available Media

Phishing websites remain a dominant enabler of cybercrime. In response, many academic AI-based phishing website detection methods have been developed, often inspiring the design of real-world systems. Although most studies report high accuracy, it remains unclear whether they meet real-world requirements such as resilience to evolving phishing tactics, robustness on diverse benign pages, interpretability, and privacy. We present PHILTER (PHishing detection literature Inspection via LLMs and Targeted Expert Review), a scalable framework for qualitatively assessing phishing website detection studies across four functionality and three security metrics. PHILTER leverages LLMs to extract evidence and draft rationales, which experts then validate and use to produce the final assessment. Applying it to 55 academic approaches reveals systemic gaps. No study fulfills all functionality and security requirements. None show evidence of effectively addressing diverse phishing tactics. Most approaches struggle to preserve privacy and adapt to evolving attacker strategies, and many risk elevated false alarms in practice due to limited testing on diverse benign pages. We also introduce a taxonomy of detection strategies (feature-based, similarity-based, identity-based, and hybrid) that highlights design trade-offs and helps explain these shortcomings. Our study reveals that accuracy-driven evaluation overlooks blind spots that undermine practical effectiveness and exposes a key open challenge: achieving high accuracy while fulfilling all functionality and security requirements. We provide actionable recommendations to guide the design of future defenses that pursue this simultaneous goal against evolving and adaptive phishing campaigns.

"Oh, what people would do with my knife?'' Navigating the Dual-Use Dilemma in PoC Exploit Development, Disclosure, and Community Dynamics

Arwa Al Alsadi and Lorenz Kustosch, TU Delft; Lamya Alowain, Independent; Michel Van Eeten and Carlos H. Gañán, TU Delft

Available Media

The cybersecurity landscape faces an escalating challenge as proof-of-concept (PoC) exploits transition from demonstrations to weaponized attacks within minutes of disclosure. While research has documented temporal dynamics and malicious deployment, a critical gap remains in understanding the human factors underlying PoC creation. Through semi-structured interviews with 16 PoC developers across diverse regions, we apply Expectancy-Value Theory to reveal PoC development as a complex motivational ecosystem where technical confidence, value assessments, and risk calculations intersect within dual-use tensions. We demonstrate that PoC development spans a continuum from crash demonstrations to weaponized exploits, shaped by multifaceted calculus rather than binary ethics. We identify three theoretical extensions: dual-use moral reasoning enabling responsibility externalization, dynamic value assessment where vendor behavior reshapes disclosure decisions, and identity navigation between ethical research and technical mastery. Vendor responsiveness, community dynamics, and legal constraints significantly influence disclosure strategies. PoC developers adopt risk-mitigation approaches when navigating tensions between security improvement and potential misuse, challenging binary conceptualizations of "responsible" versus "irresponsible" disclosure.

Raising the Flag: Detecting Missing Permission Controls in Mini-Program APIs

Zhiao Wei, University of Waterloo; Chao Wang, The Ohio State University; Haseeb-Ur-Rehman Faheem, University of Waterloo; Luyi Xing, University of Illinois Urbana-Champaign; Yousra Aafer, University of Waterloo; Zhiqiang Lin, The Ohio State University

Available Media

Mini-programs embedded within super-apps like WeChat have surged in popularity due to their flexibility and convenience, offering rich functionalities through underlying APIs. However, this tight integration introduces serious security risks: mini-programs often inherit the mobile operating system's permissions granted to their host super-app, potentially bypassing critical security checks. In this paper, we address the urgent need for more fine-grained access control tailored to mini-programs. We propose PERMSCOPE, a systematic framework to detect and analyze missing scope checks in mini-program APIs, revealing instances where Android permissions are implicitly inherited and exercised in mini-program APIs, i.e., cases where "ask and check'' primitives are absent. Our empirical analysis of four major super-apps reveals that 183 (8.85%) APIs are not properly protected at the mini-program API layer. We discuss the challenges underlying this issue and advocate for super-app developers to enforce fine-grained access control mechanisms at the API boundary, thereby strengthening the trustworthiness of the mobile super-app ecosystem.

The State of Passkeys: Studying the Adoption and Security of Passkeys on the Web

Louis Jannett, Ruhr University Bochum; Andreas Mayer and Maximilian Westers, Heilbronn University of Applied Sciences; Vladislav Mladenov, Ruhr University Bochum; Christian Mainka, University of Wuppertal; Jörg Schwenk, Ruhr University Bochum

Distinguished Paper Award Winner

Available Media

Passkeys provide a secure and phishing-resistant authentication method based on FIDO2 and WebAuthn. They have recently gained popularity, with an increasing number of websites adopting them. Nevertheless, a comprehensive security analysis that evaluates such websites at scale has not been fully addressed. We present PASSKEYS-RADAR, a continuously updated dataset that tracks the deployment of passkeys on the Internet since 2021. To build this dataset, we aggregated diverse sources, including community directories, Tranco 1M, CrUX 18M, and historic Internet archive data. We analyzed the collected data of 872 passkey-enabled websites and shed light on how passkeys are implemented and managed. We identify major differences in how websites allow users to add or delete passkeys and find that websites request authenticators to use deprecated cryptographic algorithms.

To perform a comprehensive security evaluation of passkey-enabled websites, we developed PASSKEYS-ATTACKER. The tool allows for precise manipulation of WebAuthn messages at every step of the protocol and integrates 15 attack types of which 10 were not covered in previous work. Among them, 2 attack types have critical CVSS scores. We discovered them on 18 out of 103 evaluated websites. These attacks take over user accounts, delete their passkeys, or lock them out of their accounts. Nearly half of the tested sites (53) were vulnerable to at least one attack with a high CVSS score, exposing users to threats such as phishing and session fixation.

Track 4

Software Security 4

Session Chair: Qinying Wang, EPFL

Grand Ballroom I-IV

Meerkat: Pushing the Practical Limits of Dynamic Bisection with PoC Mutation

Joseph Bursey, Christoph Sendner, and Ardalan Amiri Sani, UC Irvine; Zhiyun Qian, UC Riverside

Distinguished Artifact Award Winner

Available Media

Once a Linux kernel bug is identified, the developers perform the difficult task of patching it. Triage tools such as Syz-bisect provide a form of root cause analysis called bisection, which systematically checks historical kernel commits for the presence of the bug until it determines the Bug-Introducing Commit (BiC). Unfortunately, Syz-bisect correctly identifies the BiC for only 35% of bugs. In order to improve Syz-bisect while remaining within its time and resource constraints, we identify three areas where Syz-bisect could be improved: bug deduplication, the use of all available PoCs, and PoC mutation. Our solution Meerkat is the first dynamic bisection tool to apply scalable PoC mutation, improving over Syz-bisect by 64%. Through an in-depth manual analysis of Meerkat's bisection results, we find that dynamic bisection is critically limited by bug detectors changing with kernel versions, changes near the buggy code, and unrelated bugs that block dynamic analysis. Based on the nature of these issues, we argue that Meerkat is approaching the practical limits of what is possible with dynamic bisection in the real world. Furthermore, we correct the ground truth for 13 bugs, thus improving the dataset for future research.

BenchChecker: Assessing the Credibility of Bug-Fixing Benchmarks for LLMs

Di Wu and Xu He, George Mason University; Shu Wang, Palo Alto Networks, Inc.; Kun Sun, George Mason University

Available Media

While Large Language Models (LLMs) show promise for Automatic Program Repair (APR), existing benchmarks suffer from data contamination, in which test samples appear in training sets, artificially inflating performance. To address this, we propose BenchChecker, a framework compatible with both open-source and closed-source models that supports more credible evaluation by detecting such contamination. Grounded in the observation that persistent patches likely remain in recent repository snapshots used for training, BenchChecker employs a forensics-based two-stage detection process using only textual model outputs together with public repository history. The first stage conducts a repository presence test by probing the model with code prefixes to infer training inclusion; the second stage performs a patch presence test to verify patch persistence via similarity alignment and test-suite verification. Evaluations using StarCoder-15B, StarCoder2-15B, and CodeGen2-16B trained on The Stack v2 demonstrate that BenchChecker achieves robust detection with AUC scores around 0.9 across six benchmarks. Furthermore, applying BenchChecker to SWE-bench Verified and Multi-SWE-bench reveals that filtering contaminated samples reduces the reported resolution rates of most LLMs by more than 20% relative to the original rates on medium-difficulty tasks.

When Updates Backfire: A Black-Box Security Analysis of Desktop Software Update Mechanisms

Jie Wan, Pengcheng Xia, and Haoyu Wang, Huazhong University of Science and Technology

Available Media

Desktop software has become an essential component of modern computing, and software updates are the primary mechanism to patch vulnerabilities and deliver security fixes. However, the update process itself introduces new attack surfaces, particularly when validation of update data is incomplete or improperly enforced. Nevertheless, these risks in desktop updates are less studied because most update clients are closed-source and complex. We present UpdSight, a black-box framework that tests update security by simulating MitM attacks in realistic settings. UpdSight operates by emulating man-in-the-middle scenarios, intercepting traffic during the update process, and automatically validating the presence of critical weaknesses. By combining traffic interception, payload integrity inspection, and behavior monitoring, UpdSight provides a comprehensive assessment of update trust models across diverse software categories. We adopted UpdSight to 85 widely-used desktop applications. The results show 22 exploitable vulnerabilities, including downgrade, manifest manipulation, installer hijack, and path traversal. Among these, 16 have been confirmed by vendors. In addition, 5 CVE identifiers were assigned, covering vulnerabilities in 8 software products. Our findings highlight recurring design flaws, such as unsigned manifests and weak rollback checks, which allow attackers to gain code execution through the update channel.

VeCT: Secure and Efficient Constant-Time Code Rewriting with Vector Extensions

Qisheng Jiang and Danfeng Zhang, Duke University

Available Media

Timing channels allow attackers to extract secrets by analyzing the execution time of a victim program. Constant-time (CT) disciplines enforce security against timing attacks via data-flow/control-flow linearization (DFL/CFL). However, the rewritten constant-time code typically considerably increases the memory footprint of the original code, causing significant overhead. We present VeCT, a compiler-based code rewriter that leverages vector extensions to retain constant-time guarantees while improving performance. We first apply rigorous statistical tests to derive practical "safe-use" rules for AVX-512 instructions whose implementation details are proprietary; this analysis also reveals a previously unknown vulnerability in a state-of-the-art constant-time rewriter. Guided by these rules, VeCT introduces a novel strategy that eliminates unnecessary data loads in rewritten code, and enables vectorization to further improve efficiency. We implement VeCT based on LLVM to automatically transform code into AVX-512-based constant-time equivalents. On real-world applications like AES and Blowfish, VeCT reduces the overhead of transformed code by up to 98.9% compared to the state-of-the-art, while preserving constant-time behavior.

MULCOTAINT: Towards Efficient Multi-tag Dynamic Taint Analysis via Hardware/Software Co-design

Bing Qi, University of Chinese Academy of Sciences; Institute of Software, Chinese Academy of Sciences; Yi Yang and Xiangkun Jia, University of Chinese Academy of Sciences; Institute of Software, Chinese Academy of Sciences; Key Laboratory of System Software (Chinese Academy of Sciences); Zhengpin Qian and Huafeng Huang, University of Chinese Academy of Sciences; Institute of Software, Chinese Academy of Sciences; Purui Su, University of Chinese Academy of Sciences; Institute of Software, Chinese Academy of Sciences; Key Laboratory of System Software (Chinese Academy of Sciences)

Available Media

Multi-tag dynamic taint analysis (M-DTA) is critical in fine-grained analysis scenarios such as vulnerability analysis. However, current software solutions have serious performance problems. Although hardware solutions are promising, they are single-tag and difficult to extend to M-DTA. We propose an efficient M-DTA framework named MULCOTAINT via hardware/software co-design. We decouple the taint analysis from the normal execution with the coprocessor architecture and solve several challenges, such as designing taint calculation as vectorized calculation, managing taint tags with page tables, and providing functionality interfaces of the taint analysis engine. We build a dataset of 32 programs with 5 types and conduct the performance evaluation and vulnerability analysis experiments. The results show that MULCOTAINT has high performance and acceptable memory usage with abilities of detailed vulnerability analysis. MULCOTAINT outperforms the software solutions (TaintRabbit and PANDA) and hardware solutions (HardTaint, RAFT, and FineDIFT). The maximum difference of overhead increase based on the respective baselines could be '1.14x vs. 4409.09x' for 'MULCOTAINT vs. PANDA', while HardTaint's average overhead increase is 19.57 times that of MULCOTAINT. Although the prototype of MULCOTAINT's hardware cost is higher than embedded-oriented works RAFT and FineDIFT, it is acceptable due to M-DTA's complex logic.

Firmenstein: Scaling Dynamic Analysis for Linux-Based Firmware Services via API-Centric Intervention Code Synthesis

Yanzhong Wang, Wenhui Zhang, Ruigang Liang, Kai Chen, Yi Yang, and Zhiyu Zhang, State Key Laboratory of Cyberspace Security Defense, Institute of Information Engineering, Chinese Academy of Sciences; School of Cyber Security, University of Chinese Academy of Sciences; Junyan Jiang, School of Cyber Security, University of Chinese Academy of Sciences

Available Media

The rapid proliferation of Linux-based IoT devices necessitates rigorous security analysis. To achieve scalable, hardware-free analysis, firmware rehosting has become essential for constructing faithful emulated environments. However, existing approaches rely on inflexible heuristics to perform semantic-agnostic interventions, failing to generalize to diverse targets and leaving many firmware samples unanalyzable.

In this paper, we propose FIRMENSTEIN, a novel API-centric framework that unifies the complex and cumbersome rehosting process into a systematic program analysis and synthesis task. Specifically, we introduce an agentic cross-program analysis built upon the API Dependency Graph to automatically diagnose rehosting roadblocks. Guided by these diagnostics, FIRMENSTEIN further introduces a roadblock-driven workflow to synthesize high-fidelity intervention code that resolves environmental dependencies, thereby enabling successful firmware execution. Evaluation on an LFwC-based dataset demonstrates that FIRMENSTEIN enables 2.5×, 3.1×, and 2.1× more firmware samples to reach interact states than the state-of-the-art rehosting tools Greenhouse, FirmAE, and Penguin, respectively. Even with incomplete rehosting, FIRMENSTEIN executes 29% more basic blocks than Greenhouse. Furthermore, FIRMENSTEIN executes 5.2% more basic blocks than Greenhouse in interactive testing, validates 29 out of 31 known N-day vulnerabilities, and facilitates the discovery of 43 previously unknown vulnerabilities, with 19 CVE IDs assigned to date.

CULPA: Universal Detection of Memory-Safety Bugs in Unsafe Rust Through the Lens of Safety Requirements

Hung-Mao Chen and Bo Lu, George Mason University; Xu He, Visa Inc.; Xiaokuan Zhang and Kun Sun, George Mason University

Available Media

Rust has emerged as a leading system programming language by providing strong compile-time guarantees for memory safety. However, these guarantees do not extend to unsafe Rust, where developers may bypass compiler checks and inadvertently introduce memory-safety vulnerabilities. Although prior static analyzers have made progress in detecting such bugs, existing approaches are often fragmented: they target specific coding patterns or bug classes without modeling the underlying causes of unsafety. In this paper, we present CULPA, a universal detector for memory-safety bugs in Rust programs. The key insight of CULPA is to detect the root cause of such bugs: the violation of safety requirements in unsafe Rust contexts. To do so, CULPA first transforms the safety requirements in the standard library documentation into machine-executable predicates. Then it constructs the distinct memory segments to comply with safety requirements. Finally, CULPA collects all safety-relevant safeguards to construct the Requirement Graph. Based on the requirement graph traversal, we can determine whether violations inside unsafe blocks can be triggered. CULPA covers existing bug classes addressed by four prior static analyzers and identifies additional memory-safety vulnerabilities beyond their scope. We evaluate CULPA on the top 1,000 Rust packages. CULPA uncovers 55 previously unknown (zero-day) memory-safety bugs, 29 of which have been confirmed by developers. Most of these vulnerabilities are missed by four state-of-the-art Rust static analyzers and one LLM-based tool. To date, we have received five RustSec IDs and one CVE ID.

InSPECtor: Improving SLEIGH Specification Veracity via Proxy

Michael Chesser, Adelaide University; Paul Quirk, Defence Science and Technology Group; Douglas Cooke and Guy Farrelly, Adelaide University; Surya Nepal, CSIRO; Damith C. Ranasinghe, Adelaide University

Available Media

Processor specifications underpin critical security and program-analysis tools such as disassemblers, decompilers, and emulators, yet, their correctness is rarely examined. Errors in specifications distort program behaviour, obscure vulnerabilities, and enable analysis-evasion techniques. Validating processor specifications is a non-trivial task. Our study is a significant undertaking to enable, for the first time, the systematic validation of open-source SLEIGH language specifications, predominantly used by Ghidra. We design and implement a testing framework based on an automated oracle validation strategy by proxy. Our approach leverages the structure encoded in a specification itself to enumerate decodable instruction forms and generate targeted initial states. Then differentially test the successful decoding and emulation of those instructions by comparing emulators exercising the processor specification against hardware references.

Applying InSPECtor across diverse, open-source specifications—x86-64, AArch64, ARM/Thumb, RISC-V, MSP430—embedding differences in specification styles, author preferences, and instruction set architecture designs, we uncovered over 38,900 discrepancies that led to 125 unique bugs with proposed fixes, identifying decoding and semantic defects as well as cross-vendor inconsistencies. We distill our findings into 8 concrete recommendations to drive future improvements. Our work underscores the importance of specification correctness and provides a practical tool to substantially improve the fidelity of SLEIGH processor specifications, strengthening the reliability of downstream security and analysis tools.

Wormholes in the File System: Understanding the Misunderstanding of Symlinks

Yongheng Liu, Lei Zhang, Yuhang Zhao, and Yuzhou He, Fudan University

Available Media

Symlinks, a file system feature supported at the operating system level for more than four decades, are deeply integrated across the software stack, making them an indispensable component of modern computing environments. The symlink mechanism is highly flexible, allowing references to nearly any location in the file system and enabling complex resolution behaviors. However, this flexibility also introduces significant misunderstandings for developers, which has led to widespread misuse and become a primary vector for critical security vulnerabilities in file-handling software, e.g., arbitrary file write and code execution.

In this paper, we present the first systematic study of how symlinks are incorrectly handled and the substantial security risks. Specifically, we study how programming language standard libraries improperly utilize symlink-related system calls, and how such misuse and misunderstanding propagates to downstream applications which results in various inadequate path validations. To further understand whether these inadequate path validations could be exploited in real world, we design and implement five path validation bypass schemes and three arbitrary write methods. Our evaluation across 85 projects uncovered 16 arbitrary file write vulnerabilities, two of which have been assigned high-severity CVEs. We also demonstrate their severe impacts in real-world scenarios and propose multiple mitigation strategies. We have responsibly disclosed these issues to the affected vendors.

Track 5

Hardware Security 4

Session Chair: Eyal Ronen, Tel Aviv University

Dover Room

SoK: Deep Learning-based Physical Side-channel Analysis

Sengim Karayalcin, Leiden University; Marina Krček, Radboud University; Stjepan Picek, University of Zagreb Faculty of Electrical Engineering and Computing & Radboud University

Available Media

Deep learning-based side-channel analysis (DLSCA) represents a powerful paradigm for running side-channel attacks. DLSCA, in a state-of-the-art setting, can break multiple targets with a single attack trace, requiring minimal feature engineering. As such, DLSCA is an extremely active research area for both industry and academia. At the same time, due to domain activity, it becomes more difficult to understand the current trends and challenges. In this systematization of knowledge, we provide a critical overview of developments in DLSCA over the past years, enabling us to offer concrete suggestions and taxonomies. Moreover, we examine the reproducibility perspective and show that achieving it in full is difficult and requires a careful experimental setup.

PROBE+DETECT+MITIGATE (PDM): Enabling Cloud Tenants to Self-Defend against Microarchitectural Attacks

Arash Daneshmand, University of British Columbia Okanagan and Concordia University; Hugo Kermabon-Bobinnec, Concordia University; Lingyu Wang, University of British Columbia Okanagan and Concordia University; Makan Pourzandi, Ericsson Security Research, Ericsson Canada; Suryadipta Majumdar, Concordia University; Yosr Jarraya, Ericsson Security Research, Ericsson Canada

Available Media

Microarchitectural attacks represent a critical security concern in public cloud environments, as they can cause information leakage between cloud tenants with conflicting interests. Existing solutions usually require provider-level resources, such as hardware performance counters or host processes, which may be inaccessible to cloud tenants. The lack of awareness among cloud tenants may persuade cloud providers to postpone the deployment of vendor patches, as evidenced by patched-yet-active threats, such as PRIME+PROBE and Spectre variants. In this paper, we propose PDM, a solution that enables cloud tenants to independently detect and mitigate microarchitectural attacks without providers' help. First, PDM introduces tenant-based detection based on an interesting observation, i.e., probing the memory space of victim applications using the popular FLUSH+RELOAD attack technique can actually enable detection. Second, PDM achieves efficient tenant-based mitigation by selectively triggering obfuscation and in-memory encryption techniques upon detection. Third, we tackle several key challenges including (i) attacks not involving evictions (e.g., Spectre), (ii) the need for source code or binary instrumentation, (iii) benign noises from the victim or co-resident tenants, and (iv) the tradeoff between accuracy, delay, and overhead. Our experiments show that PDM allows tenants to detect and mitigate various microarchitectural attacks, including PRIME+PROBE and Spectre, in an accurate (e.g., ≥99.72% TPR and ≤0.13% FPR on our testbed, and ≥98.63% TPR and ≤0.83% FPR on AWS Fargate), timely (e.g., 7ms lead time for triggering mitigation), efficient (e.g., ≤2.47% overhead on SPEC CPU 2017), and robust (against both noises and evasive attacks) manner.

Charting the Cache Side-Channel Frontier: A Systematic Study of Eviction Set Construction on Apple Silicon

Han Wang, Yakun Wu, Yusi Feng, and Yinqian Zhang, Southern University of Science and Technology

Available Media

Eviction sets constitute the fundamental primitive for conflict-based cache side-channel attacks, enabling unprivileged adversaries to induce contention deterministically. However, the security landscape of Apple Silicon remains largely underexplored due to the challenges of constructing these sets. We conduct a systematic reverse engineering study on the Apple M4 to uncover the root causes of these limitations, key microarchitectural impediments: an SLC victim buffer that obscures latency-based verification, a DRAM-to-Cache Prefetcher (DCP) that violates strict LRU replacement, and an opaque SLC allocation policy modulated by L2 Insertion/Promotion Vectors (IPV). Guided by these insights, we optimize construction strategies to overcome these impediments. Our enhanced Prime-Prune-Probe (PPP) achieves 88.2% success rate for L2 eviction sets. Furthermore, we employ a hierarchical filtering strategy to construct the first practical fine-grained SLC eviction sets on Apple Silicon. Our SLC optimal approach achieves a 90% success rate, unlocking fine-grained cross-cluster attack primitives previously deemed infeasible.

TIMESLICE-SANDWICH: A GPU Side-Channel Attack Exploiting Time-Sliced Scheduling

Hodong Kim and Gyeongsup Lim, Korea University; Seunghee Shin, State University of New York at Binghamton; Youngjoo Shin and Junbeom Hur, Korea University

Available Media

Modern GPUs support resource sharing among concurrent applications, introducing the risk of side-channel attacks. While prior research has explored GPU side channels that exploit shared GPU resources, the security implications of time-sliced scheduling, a standard feature for resource sharing in today's GPUs, remain largely unexplored concerning side-channel attacks. In this study, we analyze timing variations caused by concurrent execution under the GPU's time-sliced scheduling mechanism. We begin by identifying the upper bound of a time slice and then leverage this bound to estimate the duration of a concurrent program's time slice, ultimately enabling us to infer the program's overall GPU utilization patterns.

Building on this finding, we introduce TIMESLICE-SANDWICH, a novel GPU side-channel attack that leverages variations in time-slice duration to infer and distinguish victim execution patterns. Unlike prior GPU side-channel attacks, TIMESLICE-SANDWICH does not require contention on specific shared resources. In our experiments, TIMESLICE-SANDWICH achieves an F1 score of 94.40% in neural network recovery attack; and a Top-1 accuracy of 92.84% in website fingerprinting attack on Google Chrome, both on average, demonstrating its effectiveness. Even in the presence of noise, our attack achieves an average F1 score of 73.74% for neural network recovery. Finally, we discuss potential mitigations to address side-channel risks arising from time-slice patterns in modern GPU resource-sharing architectures.

microSCALE: Static Analysis for Microarchitectural Side-channel Leakage Evaluation

Akshay Kumar E, Pranav Krishna N, Annapurna Valiveti, Pallavi Borkar, Aditi Roy, and Chester Rebeiro, IIT Madras

Available Media

Developing cryptographic implementations that are resistant to side-channel attacks is challenging in modern CPUs. Data-dependent switching activity induces subtle transitions deep in the microarchitecture that can cause sensitive information to leak through the CPU's power consumption. Existing leakage detection tools largely rely on simulation-based testing, which is not only time-consuming but is also limited to the explored execution scenarios.

We present microSCALE, a fully automated framework for detecting power side-channel leakages that occur due to transitions in the CPU's microarchitecture. By statically analyzing the CPU's RTL, microSCALE enables systematic exploration of microarchitectural leakage sources within seconds. It precisely localizes the origin of leakage in the hardware and traces it back to the corresponding software instructions. We evaluate microSCALE on several protected crypto-implementations, including GIFT, ASCON, PRESENT, and AES, running on an open-source RISC-V processor, and identify several leakage sources. For instance, we identify that a protected implementation of the GIFT cipher has about 80% of its instructions that leak on the Rocket Core. We show how microSCALE can help in hardening this implementation, reducing the side-channel leakage around 90.3%.

Spectre on RISC-V Silicon: Attacks and Defenses on Commercial Out-of-Order Processors

Lukas Gerlach, CISPA Helmholtz Center for Information Security; Marton Bognar, DistriNet, KU Leuven; Daniel Weber and Michael Schwarz, CISPA Helmholtz Center for Information Security; Jo Van Bulck, DistriNet, KU Leuven

Available Media

Speculative execution attacks have been extensively studied on mainstream x86 and ARM architectures. However, on RISC-V, research has mostly concentrated on open-source academic designs. Commercially available RISC-V silicon is widely perceived as too simple to be vulnerable, and as a result, no end-to-end attacks have been demonstrated on real hardware to date and essential software such as the Linux kernel remains unmitigated.

In this paper, we challenge that assumption. We systematically assess all commercially available out-of-order RISC-V processors (SiFive P550 and T-Head Xuantie C910/C920), finding them vulnerable to a range of Spectre attacks, and demonstrate the first Spectre attack leaking arbitrary kernel memory on real RISC-V hardware. Concerningly, our analysis reveals that mitigations in compilers, operating systems, and applications remain largely absent, and that the RISC-V instruction set lacks a dedicated speculation barrier. As a stopgap solution, we empirically characterize which instructions can halt speculation on commercial processors. We additionally audit the Linux kernel for Spectre gadgets and contribute patches, several of which have been accepted upstream. Finally, we evaluate and benchmark software-based Spectre mitigations and derive recommendations for the evolving RISC-V ecosystem, laying the groundwork for securing real hardware as it enters security-critical deployments.

UncoreBleed: AEX-Free, High-Resolution, and Low-Noise Side-Channel Attacks on SGX Enclaved Execution

Decheng Chen, South China University of Technology; Zhi Zhang, The University of Western Australia; Zhenkai Zhang, Clemson University; Xin Zhang, Shandong University; Yansong Gao, Southeast University; Yi Zou, South China University of Technology

Available Media

Trusted execution environments such as Intel SGX provide strong confidentiality and integrity guarantees by isolating enclaves from the OS and hypervisor. Prior works claim that SGX disables PMCs to mitigate side-channel attacks.

In this paper, we show that modern processors feature uncore PMCs whose behavior under SGX has not been fully evaluated. Leveraging this observation, we investigate the state of PMCs in production-mode SGX enclaves and overturn the long-held belief that performance monitoring is suppressed: uncore PMCs record events correlated with enclaved execution. We further identify a critical event in the mesh-to-memory uncore subsystem that allows address-based monitoring at 64 B granularity. Through reverse engineering, we uncover its filtering mechanism, programmability, availability, and address mapping across SGX-capable Xeon processors.

Building on the event, we present UncoreBleed, the first PMC-based, AEX-free, high-resolution, and low-noise sidechannel attack against SGX. UncoreBleed can reconstruct pictures from enclaved Libjpeg and extract RSA private keys from a single decryption, in the presence of TLBlur with AEX-Notify, the most state-of-the-art software defense on off-the-shelf SGX platforms. Our findings demonstrate that active uncore PMCs pose a previously underestimated threat to enclave confidentiality, highlighting the need to reconsider SGX's security assumptions of performance monitoring.

Behind Bars: A Side-Channel Attack on NVIDIA MIG Cache Partitioning Using Memory Barriers

Cheng Gu, University of Rochester; Reese Levine, UC Santa Cruz; Zhenkai Zhang, Clemson University; Tyler Sorensen, Microsoft and UC Santa Cruz; Yanan Guo, University of Rochester

Available Media

NVIDIA Multi-Instance GPU (MIG) is a feature designed to enable isolation and secure multi-tenancy on large data center GPUs. MIG partitions a single GPU into multiple instances, each with dedicated hardware resources such as L2 cache slices. MIG is also documented to form the foundation of NVIDIA's confidential computing stack by providing hardware-isolated trusted execution environments. However, the security claims of MIG deserve closer investigation, especially given the complexity of the GPU memory system and its many (sparsely documented) memory instructions.

In this work, we empirically examine the behavior of GPU L2 cache with MIG enabled. We find that despite the partitioning design, cross-instance L2 cache interference still occurs. Specifically, memory barriers (membars) generated in one MIG instance have side effects that propagate across L2 partitions and affect the timing of certain load operations in other instances. We also find that these membars can be triggered by specific GPU activities, such as kernel launches. Building on these observations, we develop a new timing-based side-channel attack in which an attacker in one MIG instance can infer the kernel launch patterns of a victim in another instance. We show that this attack compromises the confidentiality of widely used GPU applications, such as large language model inference, because kernel launch patterns in these applications are correlated with sensitive information.

Exploiting Hidden Resource Contention in Selective Speculation Defenses

Xiaoyu Cheng, Southeast University; Fei Tong, Southeast University; Purple Mountain Laboratories; Jiangsu Province Engineering Research Center of Security for Ubiquitous Network; Zhenyu Lei, Fang Jiang, Zhe Zhou, and Guang Cheng, Southeast University; Trevor E. Carlson, National University of Singapore (NUS)

Available Media

Transient execution attacks continue to evolve beyond cache-centric channels, motivating selective speculation defenses that aim to provide comprehensive protection with low overhead by delaying only transmit instructions. In this work, we show that several state-of-the-art selective-speculation defenses rest on shared assumptions that overlook important microarchitectural behaviors, leaving systematic blind spots that admit secret-dependent reservation-station (RS) contention.

Our analysis identifies three limitations in optimized selective-speculation designs. First, existing transmit taxonomies emphasize post-issue execution effects and miss dispatch-phase channels, such as operand-dependent μop expansion in instructions (e.g., REP-prefixed string operations) that create operand-dependent RS occupancy before execution. Second, the delay-until-resolution strategy focuses on redirect-based control leakage, but predicated instructions (e.g., x86 CMOV and RISC-V Zicond) enable secret-dependent selection without branch resolution, allowing secrets to steer operand-dependent μop expansion (e.g., REP iteration counts). Finally, the older-μop-first allocation strategy assumes unsafe contention is prevented by prioritizing non-transient μops, yet undelayed arithmetic and cache-hit memory μops can still lead to secret-dependent RS pressure through latency-amplifying dependency chains.

Guided by these findings, we construct Spectre-v1-style gadgets that bypass STT/DOLMA on x86 and RISC-V gem5 models. We further validate RS-contention effects on real CPUs, including a REP MOVSB- and REP STOSB-based proof-of-concept and a real-world RS-contention pattern identified by our LLVM pass, demonstrating realistic gadget structure and measurable signal. Finally, we propose strengthened STT mechanisms that close both existing and newly exposed gaps at moderate performance overhead.

10:30 am–11:00 am

Coffee and Tea Break

Grand Foyer North and West

11:00 am–12:30 pm

Track 1

Human Aspects 3

Session Chair: Shuang Hao, University of Texas at Dallas

Grand Ballroom VII-X

“Maybe there’s only one passkey?”: Challenges Investigating and Remediating Adversarial Passkeys

Alaa Daffalla, Cornell University; Grace Myers, Cornell Tech; Rosanna Bellini, New York University; Thomas Ristenpart, University of Toronto; Nicola Dell, Cornell Tech

Available Media

Passkeys are being actively rolled out by hundreds of web services who market them as a promising passwordless authentication method. However, it remains unclear whether people understand how to manage passkeys as part of their broader account security management, particularly in the aftermath of account compromise. We conducted a qualitative lab-based study that explores how people investigate and remediate suspicious activity when passkeys are used as a vector for illicit account access on three popular, passkey-supporting services: Google, PayPal, and LinkedIn. We recruited 31 participants with diverse technical backgrounds and tasked them with: (1) investigating a potential incident of compromise involving passkeys and (2) taking steps needed to re-secure the account.

Participants struggled to manage passkeys within account security settings and on devices, even when supported by service-provided email notifications, account security interfaces (ASIs), and streamlined wizards. The design and content of notifications and ASIs were often confusing or unclear, while the service-provided wizards were incomplete or misleading. This resulted in participants missing key steps required to discover adversarial passkeys and fully secure the account to prevent continued adversarial access. We discuss design implications and opportunities for future work to improve passkey management tools in ways that better support people experiencing account compromise.

Why Johnny Adopts Identity-Based Software Signing: A Usability Case Study of Sigstore

Kelechi G. Kalu, Sofia Okorafor, and Tanmay Singla, Purdue University; Sophie Chen, Carnegie Mellon University; Santiago Torres-Arias and James C. Davis, Purdue University

Available Media

Software signing is the most robust method for ensuring the integrity and authenticity of components in a software supply chain. Traditional signing tools burdened practitioners with key management and signer identification, creating both usability challenges and security risks. A new class of next-generation signing tools has automated many of these concerns, but little is known about their usability and its effect on adoption and effectiveness in practice. A usability evaluation can clarify the extent to which next-generation designs succeed and highlight priorities for improvement.

To fill this gap, we conducted the first usability study of Sigstore, a pioneering and widely adopted exemplar of next-generation signing. Through interviews with 17 industry experts, we examined (1) the problems and advantages associated with practitioners' tooling choices, (2) how and why their signing-tool usage has evolved over time, and (3) the contexts that cause usability concerns. Our findings illuminate the usability factors of next-generation signing tools and yield recommendations for toolmakers, adopting organizations, and the research community. Notably, components of next-generation tooling exhibit different levels of maturity and readiness for adoption, and integration flexibility is a common pain point, but potentially mitigable through plugins and APIs. Our results will help next-generation signing toolmakers further strengthen software supply chain security.

X-rated Compliance Theater: An Empirical Evaluation of European Age Verification Systems in Adult Websites

Simone Lavermicocca, Michele Carminati, and Stefano Longari, Politecnico di Milano

Available Media

Age verification is rapidly emerging as a central regulatory instrument for protecting minors online, with several jurisdictions mandating its deployment for access to adult and pornographic content. This regulatory direction raises significant privacy concerns, as it risks binding sensitive content access to identity-related attributes. It also introduces security risks, since age-verification mechanisms are often outsourced to third-party providers with limited transparency into the robustness of their verification processes. In this work, we conduct, to the best of our knowledge, the first exploratory security assessment of regulation-mandated age-verification mechanisms deployed by adult websites. Rather than treating age verification as a purely regulatory question, we empirically examine whether current deployments provide security guarantees commensurate with the privacy risks of relying on sensitive identity-related data. Our methodology combines ecosystem mapping, adversary modeling, and empirical testing across four countries, covering document-based verification, biometric age estimation, indirect signals, and website-workflow integration. Our results reveal systemic weaknesses across mechanisms and integrations under realistic threat assumptions, including failures against low-cost, widely accessible attacks. Finally, we derive concrete guidelines and design directions for mitigating the security and privacy risks exposed by current age-verification deployments.

Stayin' Alive: How Global Stolen Data Markets Thrive on Telegram

Tina Marjanov, University of Cambridge; Taro Tsuchiya, Carnegie Mellon University; Konstantinos Ioannidis and Jack Hughes, University of Cambridge; Nicolas Christin, Carnegie Mellon University; Alice Hutchings, University of Cambridge

Available Media

Stolen data acts as a catalyst for many cybercriminal activities, such as spam campaigns, spear phishing, and identity theft. Studying online communities that serve stolen data helps combat those criminal activities. While anonymous marketplaces and forums have traditionally been the primary venue for stolen data, the chat-based messaging application Telegram has emerged as a popular alternative. Given Telegram's increased accessibility to the general public, it remains unclear how stolen data communities adapt their operations to this platform, circumvent moderation efforts, and create resilient communities. In this work, we characterize: i) where stolen data communities appear within Telegram's ecosystem, ii) what types of stolen data they offer, iii) where they operate from, and iv) how they evade detection. This paper offers four main contributions. First, we provide one of the largest longitudinal datasets of Telegram stolen data channels. Over one year, we manually curate 1,521 channels and collect 14 million messages and 3.6 million shared files. We show that the stolen data communities are largely disjoint from other communities on Telegram. Second, we categorize the types of stolen data with the aim of understanding the potential cybercrime they enable. Third, while existing literature focuses on English-speaking communities, we find that many channels operate in non-English languages and source stolen data from non-English markets. Fourth, those communities deploy various techniques to evade regulation. Notably, "gateway channels" that provide links to other stolen data channels play a crucial role in increasing longevity and growth rate. We conclude by providing implications not only for academic researchers but also for Telegram and law enforcement agencies across different jurisdictions seeking to monitor and moderate those activities.

"Your imaging may be stone-cold normal, but if they look sick, they’re going to get admitted": An Investigation of Clinicians’ Perceptions of Impact & Likelihood of Security Failures

Ronald E. Thompson III and Hamza Khalid, Tufts University; Hilary Fisher, Brigham and Women's Hospital; Rhea Votipka, Beth Israel Lahey Health; Daniel Votipka, Tufts University

Available Media

Cyberattacks are a critical patient safety issue, yet security controls often fail to account for the uniqueness of the clinical environment. This paper addresses the gap in understanding clinicians' security perspectives through a mixed-methods study, with 12 interviews of US clinicians, followed by a 303-participant survey of clinicians across the US, UK, and Canada. Our findings reveal a significant misalignment between perceived threats and deployed controls. Clinicians perceive confidentiality failures (e.g., data breaches) as most likely. They view integrity failures (e.g., manipulated values) as catastrophic but trust their own expertise to ignore anomalous data. Finally, they manage likely and dangerous availability failures with analog workarounds like paper charting, introducing new risks. These results show the need to integrate clinicians into security, highlighting where existing approaches are lacking and providing recommendations for developing more effective, clinician-centered security.

Designing Wallet-Based User Intervention for Approval Phishing Mitigation

Maggie Yongqi Guan, University of Macau; Yuqi Xu, Yunlong Mao, and Wei Tong, Nanjing University; Xiaobo Zhou and Kanye Ye Wang, University of Macau

Available Media

Approval phishing is a form of Web3 phishing that exploits token approval mechanisms to trick users into granting attackers spending authority over their tokens. As attackers increasingly hijack legitimate websites, URL-based detection alone becomes insufficient, leaving crypto wallets as the last line of defense. Based on the characteristics of approval mechanisms, we propose four wallet-based interventions for mitigating approval phishing: Spending Cap Suggestion, Active Spender Warning, Passive Spender Warning, and Delayed Confirmation. We evaluate the interventions through a between-subjects experiment (n = 364) and semi-structured interviews (n = 23). Compared with the control group, the Spending Cap Suggestion condition significantly increases the likelihood that users set spending caps. The Active Spender Warning, Passive Spender Warning, and Delayed Confirmation conditions all increase cancellation rates of phishing tasks, although the increases are statistically significant only for Active Spender Warning and Delayed Confirmation conditions. The effectiveness of the interventions varies across users, as users may struggle to interpret suspicious cues and focus on transaction outcomes while overlooking approval details. Our findings highlight the need to strengthen defenses against such attacks by increasing users' awareness of post-approval consequences and supporting approval-parameter verification at the moment of authorization.

Inconsistent, Incomplete, and Insecure: A Survey of Account Security Interfaces

Arkaprabha Bhattacharya and Alaa Daffalla, Cornell University; Kevin Lee, Independent Researcher; Rosanna Bellini, New York University; Nicola Dell, Cornell Tech; Thomas Ristenpart, University of Toronto & Cornell Tech

Available Media

Despite improvements in account security, compromise remains widespread and damaging, especially when the attacker has close physical or social proximity to the victim (e.g., in terpersonal abuse settings). To help users identify unauthorized access, web services provide account security interfaces (ASIs): notifications and logs that provide information to help infer adversarial compromise. We present the largest measurement study of ASIs to date, evaluating 100 popular services.

Our study highlights an unsatisfying status quo: 29 services provided users with no way to distinguish account accesses. After categorizing ASIs using a new typology, we show that services were inconsistent in the types they deployed. Further, ASIs were often incomplete and confusing, even for expert researchers. Finally, of 61 services that offered an ASI to convey device or location descriptions, 41 (67.2%) were vulnerable to spoofing attacks that successfully obfuscate the source of the access. Based on these findings, we present six principles for improving future ASI deployments.

What You Decode Depends on Where You Stand: Distance-Based Optical QR Code

Pulkit Garg, Indian Institute of Technology Jodhpur; Robin Verma, Marshall University; Somitra Sanadhya, Indian Institute of Technology Kanpur; Gaurav Gupta, MeitY, Government of India

Available Media

Quick Response (QR) codes are ubiquitous; we can see them in authentication, payments, advertising, and other information-sharing use cases. QR code’s security relies on the assumption that every scanner decodes identical information from the same code. This work introduces Optical Quick Response (OQR) codes, a novel multi-layer QR code design that challenges this assumption. OQR uses specially designed optical patterns to encode distinctive information across distance-dependent layers within a single code, while upholding full backward compatibility with standard QR decoders. This paper presents 2-layer (OQR2) and 3-layer (OQR3) OQR configurations that achieve up to 300% higher data capacity. Exploiting this multi-layer capability, a filter evasion vulnerability is identified in which automated QR validators decode only the innermost layer of pixel-perfect images, enabling attackers to conceal malicious content in the outer layers and remain undetected. To counter this threat, a blurring-based defense is developed that recovers all layer content, enabling detection and flagging of malicious OQR codes before they reach end users. A custom dataset of 600 OQR codes was generated for evaluation. Results across six smartphones, for both printed and digital formats, demonstrate a reliable multi-layer Code Decodability Rate (CDR) of 100% for OQR2 and over 98% for OQR3. The filter evasion attack success rate against two popular QR decoding libraries reaches 100% for OQR2 and 98.67% for OQR3.

What Adults Will (and Won't) Do to Prove Their Age: Empirical Evidence from a Deceptive Web Experiment

Yanzi Lin and Cheng Zhang, Carnegie Mellon University; Madelyne Xiao, Princeton University; Lorrie Faith Cranor and Sarah Scheffler, Carnegie Mellon University

Available Media

As age verification laws proliferate across the United States and internationally, limited empirical evidence exists on how adults actually respond to these systems in practice. We conducted an IRB-approved, deceptive web experiment (n = 1,635) to examine how different age verification methods affect adults' decisions to access age-restricted content. Participants, recruited to evaluate R-rated movie clips, were randomly assigned to one of seven verification conditions: checkbox self-attestation, government-issued ID upload (with varying privacy reassurances), government-issued ID with liveness check, AI facial age estimation, or email age estimation. Completion rates varied substantially by method. Checkbox self-attestation achieved 99% completion, government-ID methods 18–27% regardless of reassurances, email 86%, and AI 51%. Follow-up surveys (n = 881) revealed low comfort even among participants who completed verification. Participants perceived government-ID and AI methods as significantly riskier, rated no method as particularly effective at preventing minors' access, and found government-ID methods significantly less convenient. Higher comfort and lower perceived risk were associated with higher completion. These findings have important implications for balancing child protection with adults' access rights. We offer recommendations for policymakers and platform operators implementing age verification systems.

Track 2

Hardware Security 5

Session Chair: Bart Preneel, KU Leuven

Grand Ballroom VI

Jailbreaking the AMD Secure Processor: Enabling Live Analysis of SEV-SNP's Undocumented Security Boundaries

Muyan Shen, Institute of Software, Chinese Academy of Sciences; School of Cryptology, University of Chinese Academy of Sciences; Hongzhan Ma, Ketong Shang, Ruofei Qu, Yu Qin, and Dengguo Feng, Institute of Software, Chinese Academy of Sciences

Available Media

AMD's Secure Nested Paging (SEV-SNP) protects virtual machines using its hardware root of trust, the AMD Secure Processor (ASP). However, its security relies on complex, opaque firmware features, such as a dynamic hot-patching mechanism, that create attack surfaces shielded from independent scrutiny. To overcome this "black-box" challenge, we present ASPBreaker, the first practical, fully deterministic jailbreak of the ASP, achieved by exploiting a novel combination of memory aliasing and time-of-check-to-time-of-use (TOCTOU) vulnerability.

Using this jailbreak as a tool for live analysis, we demonstrate how achieving arbitrary code execution on a vulnerable firmware version can be used to subvert the security of a subsequent, fully-patched one. Our analysis revealed critical flaws, enabling two practical attacks against the latest firmware: one that allows an adversary to decrypt the memory of a virtual machine and another that bypasses existing mitigations to forge attestation reports. Our findings, which were responsibly disclosed, demonstrate a fundamental break in the forward-security model of SEV-SNP and highlight the critical need for independent auditing of opaque firmware boundaries.

FABRICKED: Misconfiguring Infinity Fabric to Break AMD SEV-SNP

Benedict Schlüter, Christoph Wech, and Shweta Shinde, ETH Zurich

Available Media

Confidential computing is gaining popularity with real world cloud deployments. We present FABRICKED, a new attack that manipulates fabric routing to compromise AMD SEV-SNP. FABRICKED shows that a software adversary, by redirecting interconnect transactions, can mislead the secure co-processor to falsely initialize the system. Our primitive obtains arbitrary read and write within the victim address space, thereby violating confidentiality and integrity guarantees. FABRICKED also forges attestation reports.

Silicon Heist: (Ransom) Attacks for Cloud FPGAs via Privilege Escalation

Simon Klix, Felix Hahn, Maik Ender, Nils Albartus, and Christof Paar, Max Planck Institute for Security and Privacy (MPI-SP); Russell Tessier, University of Massachusetts

Available Media

Cloud-based FPGAs have become a billion-dollar industry, allowing users to deploy custom hardware designs with the scalability and flexibility of cloud infrastructure. Running user designs on hardware owned by the cloud service provider (CSP) introduces risks, including intentional hardware damage and Denial-of-Service (DoS) attacks against the host. To mitigate these risks, CSPs enforce security mechanisms that restrict user designs and prevent unauthorized behavior. We present a novel privilege escalation path on AMD FPGAs using (i) the Internal Configuration Access Port (ICAP) to circumvent provider defenses, (ii) incrementally escalate attacker capabilities to remote JTAG access, and (iii) investigate the resulting threat vectors.

Any typical cloud customer can maliciously acquire such ICAP access to reconfigure parts of the FPGA fabric without restrictions – re-enabling traditional cloud FPGA attacks. Through the ICAP, a user can ultimately gain remote control of the hardware's low-level JTAG interface, which enables access to the device's eFuses. This access, in turn, allows attackers to irreversibly program encryption settings, thereby disabling future reconfiguration and locking the CSPs out of their own devices. An attacker could leverage such escalated privileges for a ransomware attack in which cloud providers must pay a ransom for decryption keys to regain control of their devices – effectively introducing the first ransomware for FPGAs. Following the investigation of this novel privilege escalation path, we demonstrate its feasibility on Amazon's EC2 F1 and F2 instances and explore the impact of enabled attack vectors. We thereby expose the neglected threat of unsecured low-level hardware components in cloud environments.

STALEUS: Breaking AMD SEV-SNP via Memory Incoherence

Benedict Schlüter and Shweta Shinde, ETH Zurich

Available Media

AMD SEV-SNP provides confidential computing capabilities on AMD platforms. To secure confidential virtual machines, AMD relies on a trusted co-processor known as the Platform Security Processor (PSP). The PSP executes on an Arm core, which is connected to the x86 interconnect via dedicated hardware bridges. We show that a malicious hypervisor can configure one of these bridges to render the PSP cache-incoherent with respect to the x86 caches. We present a novel attack called STALEUS that demonstrates how to leverage this vulnerability to break SEV-SNP. We show the impact of STALEUS by achieving arbitrary read and write primitives within a fully attested CVM

XCFI: Comprehensive Control-Flow Integrity for Arm TrustZone-M

Yunju Gu, Jaeyeol Park, and Donghyun Kwon, Pusan National University

Available Media

Arm TrustZone-M (TZ-M) provides hardware-based isolation for tiny embedded systems by partitioning execution into secure and non-secure states. However, despite this isolation, memory vulnerabilities in software running within either state can still be exploited to perform control-flow hijacking attacks. To mitigate these threats, numerous control-flow integrity (CFI) studies have been proposed for embedded systems, but they have several limitations: many neglect the secure state, fail to protect control-flow events during TZ-M security state transitions, or incur prohibitive performance overhead.

In this paper, we present XCFI, a comprehensive CFI mechanism for TZ-M. XCFI employs a 32-bit control-flow identifier (CID) based protection to uniformly enforce fine-grained CFI across both secure and non-secure states, covering all control-flow events, including indirect branches, returns, and exceptions. Crucially, XCFI extends CFI enforcement to cross-state control-flow events in TZ-M, which have not been protected by prior work. Overall, XCFI provides comprehensive control-flow protection across all TZ-M execution contexts while incurring only modest runtime overhead.

StackWarp: Breaking AMD SEV-SNP Integrity via Deterministic Stack-Pointer Manipulation through the CPU's Stack Engine

Ruiyi Zhang, Tristan Hornetz, Daniel Weber, Fabian Thomas, and Michael Schwarz, CISPA Helmholtz Center for Information Security

Available Media

Confidential Virtual Machines (CVMs), such as AMD SEV-SNP, aim to protect guest operating systems from an untrusted host by encrypting state and constraining privileged control. These platforms promise isolation even in multi-tenant cloud setups where simultaneous multithreading (SMT) remains enabled. While prior attacks focus on the memory hierarchy or execution units, they largely ignore frontend configurations.

In this paper, we present StackWarp, a software-based architectural attack exploiting the stack engine on AMD Zen CPUs to modify the stack pointer within an SEV-SNP guest, fully breaking integrity. StackWarp relies on an undocumented bit within a shared model-specific register (MSR) available on AMD Zen 1–5 CPUs that enables or disables the stack engine. Our reverse engineering shows that the state of the stack engine is not correctly synchronized across the logical cores, allowing an attacker to deterministically adjust the stack pointer on the sibling logical core across Zen generations, including fully patched Zen 5. We discover StackWarp via a systematic exploration of the MSR space, including undocumented MSRs. By flipping MSR bits, we discover bits that affect SEV-SNP guests running on a sibling logical core. To demonstrate the security impact, we show StackWarp in four end-to-end attacks on SEV-SNP guests: RSA-CRT private-key recovery, OpenSSH password-authentication bypass, and privilege escalations using either sudo or a kernel-mode ROP chain. We conclude with software hardening guidance and argue for a microcode or hardware change that prevents cross-core control of the stack engine when CVMs are active. Our results show that leaving SMT enabled undermines SEV-SNP integrity guarantees today.

M-Step: A Single-Stepping Framework for Side-Channel Analysis on TrustZone-M

Cristiano Rodrigues, Centro ALGORITMI, Universidade do Minho; Marton Bognar, DistriNet, KU Leuven; Sandro Pinto, Centro ALGORITMI, Universidade do Minho; Jo Van Bulck, DistriNet, KU Leuven

Available Media

Trusted Execution Environments (TEEs) have become a key technology for isolating sensitive enclave applications from untrusted operating systems. Extensive research on high-end platforms like Intel SGX and TDX, AMD SEV, and Arm TrustZone-A has exposed their limitations in terms of software-based side-channel analysis, amplified by specialized single-stepping attack frameworks that exploit privileged timer interrupts to execute enclaves one instruction at a time. Meanwhile, TEEs are increasingly deployed on resource-constrained IoT devices, with Arm TrustZone-M emerging as a leading solution, which, however, remains largely unexplored for high-resolution, software-based side channels.

This paper presents M-Step, an open and extensible single-stepping attack framework for TrustZone-M. While Cortex-M microcontrollers feature precise timers and deterministic behavior, achieving precise, instruction-level stepping remains challenging due to (i) the absence of virtual memory and page tables used in high-end frameworks; and (ii) Cortex-M's unique interrupt behavior, where certain multi-cycle instructions are abandoned or paused to reduce latency. To overcome these challenges, we extensively profile interrupt handling CPU behavior and develop a novel approach that uses previously dismissed interrupt-latency leakage to dynamically adjust the timer interrupts. We demonstrate M-Step's improved resolution and practicality by discovering previously unknown vulnerabilities in the latest Arm Mbed TLS library that enable single-trace, deterministic attacks recovering full RSA keys from a TrustZone enclave.

Memclave: Secure In-Memory Enclave for Untrusted Hosts

Amit Choudhari, CISPA Helmholtz Center for Information Security; Fabian van Rissenbeck, Technische Universität Dortmund; Christian Rossow, CISPA Helmholtz Center for Information Security

Available Media

Cloud platforms run data-intensive workloads in multi-tenant settings, where frequent CPU–memory traffic can leak access patterns via cache side channels. Processing-in-Memory (PIM) devices such as UPMEM move computation into DRAM, sharply reducing data movement and shrinking the CPU cache footprint. However, commercial PIM architectures expose a host-programmed control plane and host-shared module memory, leaving device-resident code and data vulnerable to a compromised host. Existing secure-PIM proposals either add encryption/access-control hardware or rely on heavyweight host-side cryptographic protocols, complicating practical deployment.

We present Memclave, a software-only framework that brings code integrity and data confidentiality to commodity PIM without hardware changes. A TPM-attested hypervisor permanently isolates the PIM's control plane from host access at boot. On each in-memory core, a trusted loader authenticates the user kernel and establishes a per-session protected data path. Memclave preserves the programming model and kernel code: host applications replace a small set of data-movement calls with secure drop-ins, keeping the trusted computing base small and porting effort low. We implement Memclave on off-the-shelf UPMEM DIMMs and evaluate it across the PrIM benchmark suite, covering heterogeneous memory-access, compute, and synchronization patterns. After a one-time  100ms authenticated load, in-memory kernel time remains close to the PIM baseline: Multilayer Perceptron (MLP) stays within 1.5× at practical sizes, and First Search (BFS) is 1.1× on some graphs with modest rise as number of frontier levels increase.

Track 3

Applications of Cryptography 6

Session Chair: Dimitris Kolonelos, University of California, Berkeley

Grand Ballroom V

Scribe: Low-memory SNARKs via Read-Write Streaming

Anubhav Baweja, Pratyush Mishra, Tushar Mopuri, Karan Newatia, and Steve Wang, University of Pennsylvania

Available Media

Succinct non-interactive arguments of knowledge (SNARKs) enable a prover to produce a short and efficiently verifiable proof of the validity of an arbitrary NP statement. Recent constructions of efficient SNARKs have led to interest in using them for a wide range of applications, but unfortunately, deployment of SNARKs in these applications faces a key bottleneck: SNARK provers require a prohibitive amount of time and memory to generate proofs for even moderately large statements. While there has been progress in reducing prover time, prover memory remains an issue.

In this work, we describe Scribe, a new low-memory SNARK that can efficiently prove large statements even on cheap consumer devices such as smartphones by leveraging a plentiful, but heretofore unutilized, resource: disk storage. Instead of storing its (large) intermediate state in RAM, Scribe's prover instead stores it on disk. To ensure that accesses to state are efficient, we design Scribe's prover in a read-write streaming model of computation that allows the prover to read and modify its state only in a streaming manner.

We implement and evaluate Scribe's prover, and show that, on commodity hardware, it can easily scale to circuits with 228 gates while using less than 750MB of memory and incurring only minimal proving latency overhead (10%) compared to a state-of-the-art memory-intensive baseline (HyperPlonk [EUROCRYPT 2023]) that requires much more memory.

Rarus: A Succinct and Efficient Range Proof for Polynomial-based Vector Commitment

Xinyang Yang, Wenjie Qu, Yanpei Guo, and Jiaheng Zhang, National University of Singapore

Available Media

Range proofs enable a prover to convince a verifier that a committed value lies within a specific interval without revealing additional information. They are fundamental to privacy-preserving systems including anonymous credentials, e-voting, e-cash, and cryptocurrencies like Monero and Grin. A critical challenge is efficiently proving that multiple committed values simultaneously satisfy range constraints while minimizing communication overhead.

Vector commitment schemes provide a promising approach to this problem. Missileproof (CCS'24) recently proposed a range proof for vector commitments achieving O(1) proof size and verifier time, but with prover complexity of O(Nℒlog(Nℒ)) for proving ℒ values in [0,2N).

We present Rarus, an efficient range proof for polynomial-based vector commitments that achieves optimal asymptotic complexity across all metrics. Our key innovation is replacing binary decomposition with optimized b b-ary decomposition, coupled with Bi-variate Zero-Test and accelerated Uni-variate Sum-Check protocols. Rarus achieves O(1) proof size and verifier time, while reducing prover time to O (Nℒ / log (Nℒ)) G + O (Nℒ)F, where G and F denote group and field operations respectively. In addition , our protocol supports arbitrary ranges [0,R) beyond powers of two. Experimental results demonstrate that Rarus achieves a 20× speedup over both Bulletproofs and Missileproof when proving 16,384 values in [0,264 =).

Garuda and Pari: Faster and Smaller SNARKs via Equifficient Polynomial Commitments

Michel Dellepere, Independent; Pratyush Mishra and Alireza Shirzad, University of Pennsylvania

Available Media

SNARKs are powerful cryptographic primitives that allow a prover to produce a succinct proof of a computation. Two key goals of SNARK research are to minimize the size of the proof and to minimize the time required to generate it. In this work, we present new SNARK constructions that push the frontier on both of these goals.

Our first construction, Pari, is a SNARK that achieves the smallest proof size amongst all known SNARKs. Specifically, Pari achieves a proof size of just two group elements and two field elements, which, when instantiated with the BLS12-381 curve, totals just 160 bytes. This is smaller than the sizes for Groth16 [Groth, EUROCRYPT '16] and Polymath [Lipmaa, CRYPTO '24]. Pari also achieves the lowest known gas cost for on-chain SNARK verification, reducing the gas cost by 6% compared to Groth16 and 17% compared to FFLONK.

Our second construction, Garuda, is a SNARK that reduces proof generation time by supporting, for the first time, arbitrary "custom" gates and free linear gates (in terms of cryptographic costs). These benefits enable significant prover-time savings compared to state-of-the-art SNARKs.

Both constructions rely on a new cryptographic primitive: "equifficient" polynomial commitment (EPC) schemes that enforce that committed polynomials have the same representation in particular bases. We provide both rigorous security definitions for this primitive as well as efficient constructions for univariate and multilinear polynomials.

Our constructions are obtained via a new compiler that obtains a succinct argument by combining polynomial IOPs with our EPC schemes.

Prezta: Provable Remote Execution of Zero-Trust Authorization using SNARKs

Zhongjing Wei, Osaid Muhammad Ameer, Yupeng Zhang, and Nikita Borisov, University of Illinois Urbana-Champaign

Available Media

Modernizing the security of operational technology systems that control critical infrastructure has become a pressing challenge. Because edge devices have limited capabilities, modernization has relied on application gateways that interface with identity management systems and enforce access policies. These gateways are powerful enough to perform complex authorization decisions and support zero-trust architectures but create major deployment and management burdens: they must be collocated with remote, distributed edge devices, kept up to date with security patches, and managed with minimal downtime.

We propose Provable Remote Execution of Zero-Trust Authorization (Prezta), an architecture that eliminates these gateways by evaluating policies within a zero-knowledge virtual machine (zkVM) running on the client. The zkVM produces a succinct proof of authorization that edge devices can verify efficiently, extending the zero-trust security envelope to the edge. Policies and identity management schemes can evolve without updating edge devices.

To demonstrate the feasibility of Prezta, we implement a prototype, built using the RISC Zero zkVM, that supports XACML 3.0 policies and JWT identity claims. While zkVMs introduce substantial proof overhead, we mitigate this by compiling policies to Rust code and pre-compiling regular expressions. Combined with optimized signature verification and JWT parsing, these measures reduce prover time by more than an order of magnitude. Our compiler correctly implements 83% of the XACML 3.0 conformance suite, with proof generation completing in tens of seconds on a desktop. Verification, by contrast, takes only tens of milliseconds—fast enough for even resource-constrained edge devices.

Efficiently Provable Approximations for Non-Polynomial Functions

Sriram Sridhar, University of California, Berkeley; Shravan Srinivasan, Lagrange Labs; Dimitrios Papadopoulos, Hong Kong University of Science and Technology; Charalampos Papamanthou, Lagrange Labs and Yale University

Available Media

Zero-Knowledge Proofs (ZKPs) are now widely used to verify the correctness of various types of computations. However, despite phenomenal advancements, current ZKPs are inefficient for applications that need accurate evaluation of non-polynomial functions over floating-point numbers, such as machine learning, decentralized finance, scientific computing, and geolocation. Current state-of-the-art approaches typically emulate floating-point numbers using fixed-point representations (via quantization), and handle non-polynomial functions using lookup tables, piece-wise or low-degree polynomial approximations, which lead to sub-optimal performance and/or loss in accuracy or generality, limiting their potential for adoption in practice.

In this work, we present a general framework for approximating a large class of non-polynomial functions using Gauss-Legendre quadrature, which supports efficient ZKPs of correct computation. We show that our approach can scale to decrease the error up to the inherent limits imposed by quantization, without increasing the multiplicative circuit depth beyond a small constant (≤ 4). This is a strong deviation from prior approximation techniques, where decreasing the error leads to increased multiplicative depth– the main factor determining the error growth of an approximation. We implement and evaluate our approach in Noir/Barretenberg, and we obtain absolute errors 2−256× lower than comparable baselines for most non-polynomial functions with low prover overhead. We also demonstrate an efficient prover and low errors for high-accuracy applications in DeFi and astronomy that require non-polynomial functions, again obtaining errors 4−64× lower than the baseline approximations.

TensorZKP: Repurposing GPU Tensor Cores for High-Performance Zero-Knowledge Proofs

Tao Lu, Jipeng Zhang, and Yanpei Guo, National University of Singapore; Xuanming Liu, Zhejiang University; Wenjie Qu, National University of Singapore; Zonghui Wang and Wenzhi Chen, Zhejiang University; Jiaheng Zhang, National University of Singapore

Available Media

GPU Tensor Cores, specialized hardware units designed to accelerate matrix multiplication, have served as the primary engine behind the AI revolution. Given the exponential performance gains they have delivered, aligning cryptographic implementations with this hardware evolution is critical. This is particularly acute for zero-knowledge proofs (ZKPs), a cryptographic primitive that currently grapples with high proof generation costs. Existing GPU implementations for ZKPs rely exclusively on general-purpose SIMT cores, leaving the massive computational power of Tensor Cores untapped.

In this paper, we introduce TensorZKP, the first GPU framework to harness Tensor Cores for ZKP acceleration. Since Tensor Cores are designed for low-precision matrix multiplication, mapping ZKP's arithmetic to this hardware is non-trivial. To bridge this gap, we develop Tensor-Core-compatible finite field arithmetic and reformulate ZKP modules, specifically sum-check protocols and Spielman code, into matrix multiplication tasks. Furthermore, we design an asynchronous warp-specialized framework that pipelines memory access, Tensor Core matrix operations, and SIMT-based modular reductions. We instantiate these optimizations with HyperPlonk as the Polynomial Interactive Oracle Proof (PIOP) and Brakedown as the Polynomial Commitment Scheme (PCS) to enable end-to-end proof generation.

The evaluation results show that TensorZKP exhibits remarkable efficiency. At a 2^25 scale, the underlying building blocks complete in 0.85 ms for inner product, 0.91 ms for scalar-vector multiplication, 4.04 ms for degree-2 sum-check, and 11.58 ms for the encoder. For a circuit with 2^25 multiplication gates, TensorZKP achieves a proof generation time of only 215.28 milliseconds, representing a 955× speedup over the CPU baseline and a 36.2× improvement over state-of-the-art SIMT-based GPU implementations.

Icefish: Practical zk-SNARKs for Verifiable Genomics

Alexander Frolov, Maurice Shih, Rob Patro, and Ian Miers, University of Maryland

Available Media

Individual genomic data is a uniquely sensitive type of user data. While many papers have considered using Multi-Party Computation (MPC) or Fully Homomorphic Encryption (FHE) to allow collaborators to study combined genomic datasets they cannot share, few have considered verifying the results of genomic computations, either in research studies or in the emerging area of personalized genetic therapies.

In this paper, we initiate the first systematic study of zero-knowledge proofs for verifiable genomics, providing both building blocks for verifying common operations in computational genomics, such as sequence alignment, and exploring two end-to-end applications:

Verifiable Genome-Wide Association Studies: A Genome-Wide Association Study (GWAS) study operates over a repository of genomic data, identifying statistical correlations between genetic variations and observed traits or medical conditions. Our system enables third parties to verify that research was honestly computed over an authenticated, untampered database, ensuring both the integrity of the underlying data set and the correctness of the resulting science. We achieve practical performance (<40 minutes proving time) for studies of sizes equal to those in the existing genomics literature.

Verifiable CRISPR eligibility: We propose using zk-SNARKs in the context of gene engineering (e.g. CRISPR). To our knowledge, this is a new use case for zk-SNARKs. We implement and optimize models for detecting "on-target" and "off-target" sites for a CRISPR probe in zk-SNARKs, so users can, for example, demonstrate eligibility for a therapy or trial without having to reveal their own DNA sequence.

In support of these applications, we develop new building blocks, like zero-knowledge proofs of sequence alignment that are 30x faster than the prior state of the art, and storage-efficient indexes for Merkle trees for large scale genomic data that asymptotically reduce storage costs.

Shred-to-Shine Metamorphosis of (Distributed) Polynomial Commitments

Weihan Li, School of Cyber Science and Technology, Beihang University; Ant Group; Zongyang Zhang, School of Cyber Science and Technology, Beihang University; Sherman S. M. Chow, The Chinese University of Hong Kong; Yanpei Guo, National University of Singapore; Boyuan Gao, School of Cyber Science and Technology, Beihang University; Xuyang Song, Anoma; Yi Deng, School of Cryptology, Xidian University; Jianwei Liu, School of Cyber Science and Technology, Beihang University

Available Media

Succinct non-interactive arguments of knowledge (SNARKs) rely on polynomial commitment schemes (PCSs) to verify polynomial evaluations succinctly. High-performance multilinear PCSs (MLPCSs) from linear codes reduce prover cost, and distributed MLPCSs cut it further by parallelizing commitment and opening across provers. Employing a fast Reed–Solomon interactive oracle proof of proximity (FRI), we propose PIPFRI, an MLPCS that combines the linear-time proving of linear-time-encodable-code PCSs with the compact proofs and fast verification of Reed–Solomon (RS) PCSs. Reducing fast Fourier transform and hash overhead, PIPFRI is 10× faster to prove than the RS-based DeepFold (USENIX Security '25) while keeping competitive proof size and verifier time. Measured against Orion (CRYPTO '22) from linear-time-encodable codes, PIPFRI proves 3.5× faster and reduce proof size and verifier time by 15×. As a linearly scalable distributed variant, we propose DEPIPFRI, which adds accountability and distributes a single polynomial across provers, enabling the first code-based distributed SNARK for general circuits. Notably, compared with DeVirgo (CCS '22), which lacks accountability and supports only multiple independent polynomials, DEPIPFRI improves prover time by 25× and inter-prover communication by 7×. We identify shred-to-shine as the key insight: partitioning a polynomial into independently handled fragments while maintaining proof size and verifier time. Hitting the pairing regime, this insight yields a group-based MLPCS with a 16× shorter structured reference string (SRS) and a 10× faster opening time than a multilinear variant of Kate–Zaverucha–Goldberg (TCC '13).

Boosting Efficiency and Security in Arithmetization-Oriented Hashing for Zero-Knowledge Proof Systems

Elena Andreeva, TU Wien; Rishiraj Bhattacharyya, University of Birmingham; Arnab Roy, University of Innsbruck; Stefano Trevisani, TU Wien

Available Media

Cryptographic compression functions are a core component of vector commitment schemes, including Merkle tree commitments, which are widely used in modern ZK-SNARK and STARK frameworks. Arithmetization-Oriented (AO) compression functions minimize multiplicative complexity over the framework's native field F_p, making them significantly more efficient than bit-oriented designs in algebraic circuits. To date, AO compression functions have been almost exclusively constructed by applying the Sponge mode to an AO permutation.

In this work, we introduce two novel approaches for building permutation-based AO compression modes: the PA family, based on a Permutation with feedforward Addition, and PAX, as an eXtension of the PA family. We formally establish that, in contrast to the Sponge construction, our modes achieve optimal collision and preimage resistance. We also prove that PAX is indifferentiable from a random oracle, further strengthening its security and composability guarantees. We further show that variable-input-length hash functions can be safely instantiated from the PA(X) modes by applying appropriate domain extenders.

Beyond their strong security guarantees, our modes provide a framework that unifies and extends the description of several recently proposed modes that have been studied via cryptanalysis but do not come with provable security guarantees, including Jive and Trunc, as used in the AO designs Anemoi and Poseidon2.

Finally, through extensive experimental evaluation, we compare the concrete efficiency improvement that our modes offer compared to the Sponge approach over two popular AO permutation designs, Poseidon permutation and Rescue. For 128 bits of collision resistance, our modes can achieve up to a 2x speed-up over Sponge for equivalent compression rates in a software implementation. When considering R1CS arithmetization in the Groth16 framework, the PA(X) preimage-verification circuit can be 10% faster than Sponge. In the Plonky2 framework, PA(X) can achieve up to a 60% speed-up.

Track 4

Security of ML 6

Session Chair: Fnu Suya, University of Tennessee

Grand Ballroom I-IV

Sy-FAR: Symmetry-based Fair Adversarial Robustness

Haneen Najjar, Eyal Ronen, and Mahmood Sharif, Tel Aviv University

Available Media

Security-critical machine-learning (ML) systems, such as face-recognition systems, are susceptible to adversarial examples, including real-world physically realizable attacks. Various means to boost ML's adversarial robustness have been proposed; however, they typically induce unfair robustness: It is often easier to attack from certain classes (e.g., individuals) or groups (e.g., genders) than from others. Several techniques have been developed to improve adversarial robustness while seeking perfect fairness between classes. Yet, prior work has focused on settings where security and fairness are less critical (e.g., classifying objects such as cars and ships).

Our insight is that achieving perfect parity in realistic fairness-critical tasks, such as face recognition, is often infeasible—some classes (e.g., siblings) may be highly similar, leading to more misclassifications between them. Instead, we suggest that seeking symmetry—i.e., attacks from class i to j would be as successful as from j to i—is more tractable. Intuitively, symmetry is desirable because class resemblance is a symmetric relation in most domains. Additionally, as we prove theoretically, symmetry between individuals induces symmetry between any set of sub-groups, in contrast to other fairness notions where group-fairness is often elusive.

We develop Sy-FAR, a technique to encourage symmetry while also optimizing adversarial robustness and extensively evaluate it using five datasets, with three model architectures, including against targeted and untargeted realistic attacks. The results show Sy-FAR significantly improves fair adversarial robustness compared to state-of-the-art methods. Moreover, we find that Sy-FAR is faster and more consistent across runs. Notably, Sy-FAR also ameliorates another type of unfairness we discover in this work—target classes that adversarial examples are likely to be classified into become significantly less vulnerable after inducing symmetry.

The Prompt Stealing Fallacy: Rethinking Metrics, Attacks, and Defenses

Zehang Deng, Swinburne University of Technology and CSIRO's Data61; Haoyang Li, The Hong Kong Polytechnic University; Wanlun Ma, Swinburne University of Technology​; Ruoxi Sun and Derui Wang, CSIRO's Data61; Minhui Xue, CSIRO's Data61 and Responsible AI Research (RAIR) Centre, Adelaide University; Haibo Hu, The Hong Kong Polytechnic University; Sheng Wen and Yang Xiang, Swinburne University of Technology

Available Media

Text-to-image (T2I) models are increasingly embedded in creative workflows, where well-crafted prompts function as valuable forms of intellectual property (IP). However, these models are susceptible to prompt stealing attacks (PSAs), where adversaries aim to reconstruct the original prompts used to generate images. In this paper, 1) we identify key shortcomings in current evaluation practices and propose two improved metrics: Style Similarity (SS) and a novel Prompt Significance (PS) score, which together provide a more faithful assessment of PSA effectiveness. Rather than existing metrics that rely solely on semantic similarity between original and stolen information across text or image modalities, the new metrics PS and SS assess attack effectiveness with a more practical focus by explicitly accounting for the importance of modifiers and the style replication of images generated from stolen prompts. 2) Through extensive evaluation using these metrics, we find that existing PSA methods, ranging from soft prompt stealing in white-box settings to hard prompt stealing in black-box settings, are not as effective as reported, especially in recovering high-contribution prompt components. We attribute this to fundamental constrains: white-box methods suffer from mismatched optimization objectives that poorly align with token-level visual semantics, while black-box approaches experience severe information loss due to their decoupling from the target T2I model's generation process. 3) We further introduce PromptThief, a black-box PSA framework that addresses the information loss in prior methods by leveraging reinforcement learning with STS and SS to guide high token-level contribution recovery. PromptThief significantly outperforms existing baselines across multiple metrics and real-world scenarios. 4) We propose and evaluate two defense mechanisms: an adversarial-example-based active approach and a passive scheme through feature-level prompt watermarking. Our evaluation reveals that the active defense offers only limited robustness against adaptive PSAs, highlighting the need for further exploration in this direction. In contrast, the passive watermarking scheme demonstrates strong and consistent detection performance, even under various image transformations, offering a practical and reliable path forward for prompt IP protection.

Quantifying Large Language Model Attacks Through the Lens of Model Cognition

Xiuming Liu, Chaoxiang He, Xuanran Yu, Jichen Chai, Feiyue Xu, Sheng Hang, and Hanqing Hu, Shanghai Jiao Tong University; Bin Benjamin Zhu, Microsoft Corporation; Hongsheng Hu, Shi-Feng Sun, Dawu Gu, and Shuo Wang, Shanghai Jiao Tong University

Available Media

Large language models (LLMs) are vulnerable to malicious inputs that elicit harmful content. Current safety mechanisms, such as keyword filters or output moderation, largely ignore internal model dynamics. We show that safety-relevant features correlated with harmful prompting are strongly separable under lightweight probes in intermediate hidden states (up to 99% accuracy) before generation, revealing that such features persist internally even when models produce compliant outputs. Leveraging this observation, we introduce layer-wise toxicity probes and a multi-layer complementary detection framework that fuses signals from diverse depths. Our lightweight Sentinel (<5M parameters) halves false negatives compared to generation-level refusal and maintains over 94% detection accuracy under adversarial attacks—where baselines drop by 32%. Sentinel also outperforms Llama-Guard-3-8B on heterogeneous harmful prompting across seven open-weight LLMs (1.5B→72B) and multiple benchmarks (I2P, SneakyPrompt, MMA, Labelled, PIJ, ChatAlpaca, and Multi-turn Jailbreak). Beyond detection, our method provides the first quantitative, layer-resolved map of how safety-relevant signals emerge, propagate, and degrade within LLMs, enabling interpretable, inside-out alignment and diagnostics. This paper contains potentially sensitive and offensive content, including but not limited to NSFW material, hate speech, discrimination, and other harmful text. Reader discretion is advised.

SoK: Colluding Adversaries in Machine Learning Pipelines

Vasisht Duddu, Lipeng He, and Asim Waheed, University of Waterloo; N. Asokan, University of Waterloo and KTH Royal Institute of Technology

Available Media

Machine learning (ML) models are susceptible to various security, privacy, and fairness risks. Adversaries with different characteristics (i.e., objectives, knowledge, and capabilities) can collude by executing one attack to amplify others. Existing work lacks a systematic framework to explore collusion among adversaries, and to study the implications of the adversaries' characteristics. We present a framework covering collusion (a) between train- and inference-time adversaries, and (b) among inference-time adversaries. Our framework accounts for factors enabling collusion between adversaries. We propose a guideline to conjecture about the potential for collusion using enabling factors. We use it to explain prior work, conjecture about unexplored collusions, and empirically validate five such cases. Finally, we discuss how adversaries' characteristics influence the potential for collusion.

Certified in Theory, Broken in Practice: Assumption Gaps in Cryptographic Model Certification

Carter Luck, University of Massachusetts Amherst; Olive Franzese-McLaughlin, Vector Institute and University of Toronto; Elisaweta Masserova, Carnegie Mellon University; Akira Takahashi and Antigoni Polychroniadou, J.P. Morgan AI Research and AlgoCRYPT CoE; Nicolas Papernot, Vector Institute and University of Toronto

Available Media

Privacy-preserving machine learning auditing protocols allow auditors to assess models for properties such as accuracy or fairness, without revealing their internals or training data. This makes them especially attractive for auditing models deployed in sensitive domains such as healthcare or finance. For these protocols to be meaningful in real-world audit settings, though, their guarantees must reflect how the model will behave once deployed, rather than merely certifying its behavior during an audit. Existing security definitions often miss this mark: most certify model behavior only on a fixed audit dataset, without ensuring that the same guarantees generalize to other datasets drawn from the same distribution. We show that a model provider can attack many cryptographic model certification (CMC) schemes built on secure zero knowledge proofs (ZKP) by carefully engineering training data, resulting in models that exhibit benign behavior during an audit, but pathological behavior in practice. For example, we empirically demonstrate that an attacker can certify that a model achieves over 99% accuracy on an audit dataset, but less than 30% accuracy on fresh samples from the same distribution.

To address this gap, we formalize rigorous cryptographic security notions tailored to CMC frameworks, introduce a generic protocol template, and prove that it satisfies these requirements. Our results thus offer both cautionary evidence about existing approaches and constructive guidance for designing secure, privacy-preserving ML auditing protocols.

MATE: Policy-Aware Security Auditing for Mobile Agents via Synthesis-Driven Trajectory Learning

Changyue Jiang, Fudan University and Shanghai Innovation Institute; Jiayi Wang, Xin Wen, Jiarun Dai, and Geng Hong, Fudan University; Xudong Pan, Fudan University and Shanghai Innovation Institute

Available Media

Mobile agents powered by foundation models now automate complex, multi-step workflows on real devices, but their trajectories can violate app-specific security policies. Existing trajectory-level defenses rely on LLM prompting or rigid rules, and thus fail to support fine-grained, natural-language policies that generalize across apps and tasks. In this work, we introduce MATE, a lightweight, policy-conditioned auditor that encodes both agent trajectories and natural-language security policies to determine whether a trajectory violates a given policy and to explain why. Treating policies as editable text rather than fixed model parameters allows MATE to handle user-defined and evolving requirements without retraining. To construct MATE, we build a knowledge base by extracting app descriptions, workflows, and policies from hundreds of popular mobile apps worldwide, and synthesizing over 140K semantically realistic, policy-conditioned trajectories with a multi-stage pipeline. We further release MATEBench, a trajectory-level auditing benchmark with two synthetic subsets and one real-world subset of manually collected trajectories. Models trained with our synthesis-driven trajectory learning achieve over 95% accuracy on MATEBench, retain strong performance on external safety benchmarks, and audit trajectories from Zhipu's AutoGLM and Alibaba's Mobile-Agent on real devices with over 95% accuracy, outperforming prior methods by over 20%. MATE shows that practical, fine-grained security auditing for heterogeneous mobile agents is both feasible and effective.

SoK: Attack and Defense Landscape of Agentic AI Systems

Juhee Kim, UC Berkeley and Seoul National University; Wenbo Guo, UC Santa Barbara; Dawn Song, UC Berkeley

Available Media

AI agents that integrate large language models with non-AI tool components are rapidly emerging in real-world applications, offering unprecedented automation and flexibility. However, this flexibility introduces complex security challenges that differ from traditional software systems. In this paper, we present the first comprehensive systematization of knowledge on AI agent security, analyzing the design space, attack landscape, and defense mechanisms for secure AI agent systems. In addition, we identify open challenges for future research in this emerging domain. Our work provides the first systematic framework for understanding AI agent security risks and defense strategies, serving as a foundation for building secure agentic systems and advancing research in this critical area.

Trusting What You Cannot See: Auditable Fine-Tuning and Inference for Proprietary AI

Heng Jin, Chaoyu Zhang, and Hexuan Yu, Virginia Tech; Shanghao Shi and Ning Zhang, Washington University in St. Louis; Y. Thomas Hou and Wenjing Lou, Virginia Tech

Available Media

Cloud-based infrastructure has become the dominant platform for deploying large models, particularly large language models (LLMs). Fine-tuning and inference are increasingly delegated to cloud providers for simplified deployment and access to proprietary models, yet this creates a fundamental trust gap. Although cryptographic and TEE-based verification approaches exist, prohibitive proving costs and limited TEE memory prevent them from scaling to modern LLMs, leaving clients unable to practically audit these processes. This lack of transparency creates concrete security risks that can silently compromise service integrity.

We present AFTUNE, an auditable and verifiable framework that ensures the computational integrity of cloud-based fine-tuning and inference. AFTUNE incorporates a lightweight recording and spot-check mechanism that produces verifiable traces of execution. These traces enable clients to later audit whether the fine-tuning and inference processes followed the agreed configurations, by verifying sampled execution blocks inside a TEE, each covering only a small portion of the model and the execution trace. Our evaluation shows that AFTUNE adds modest overhead and makes auditing practical for clients.

Certification of Machine Learning Models via Directional Sharpness

Gefei Tan, Northwestern University & Google; Adrià Gascón, Google; Sarah Meiklejohn, University College London; Mariana Raykova, Google

Available Media

In machine learning, model certification has been identified as an important method for gaining assurance about a model's trustworthiness and quality. A model's quality is largely determined by its ability to generalize, i.e., to perform well on data beyond what it was trained on. It is not possible to certify generalization directly, however, as it depends on unknown data and is not directly measurable. Proxies such as test accuracy can be misleading when the training process is perturbed (intentionally or accidentally), and metrics such as sharpness—which has an empirically supported link to generalization—are computationally expensive and can also serve as unreliable signals when training deviates from a prescribed procedure. In this work, we propose directional sharpness, a metric designed to efficiently and reliably indicate generalization despite potential training deviations. We provide empirical and analytical evidence that directional sharpness (1) correlates more strongly with generalization than existing metrics and (2) identifies models with poor generalization more reliably than existing metrics. Furthermore, directional sharpness is efficiently computable in model auditing settings, where the verifier has access to training data, and via zero-knowledge proofs that certify quality without revealing training data.

Track 5

Network Security 4

Session Chair: Eric Pauley, Virginia Tech and Terrace Networks

Dover Room

When Address Learning Goes Wrong: Inducing Forwarding Loops and DoS Amplification in SDN

Dezhang Kong, Zhejiang University; Yilun Zhang, Purdue University; Zekun Xie, Xi'an University of Posts and Telecommunications; Ningpeng Zheng, Yancheng Institute of Technology; Shi Lin, Zhejiang Gongshang University; Zhenhua Xu, Zhejiang University; Minghao Li, Chongqing University; Zhebo Wang, Xiang Chen, and Changting Lin, Zhejiang University; Dong Zhang, Fuzhou University; Xuan Liu, Yangzhou University; Chunming Wu and Meng Han, Zhejiang University

Available Media

Address learning is a fundamental SDN service that maintains a dynamic mapping from host addresses to switch ports, supporting many critical network applications. However, the address learning's core position makes it an attractive attack target. Meanwhile, unfortunately, it lacks security protection from a global view, resulting in the fact that individual, normal events can collectively cause damage. Based on it, this paper proposes LoopGen, a new attack targeting the address learning mechanism. LoopGen shows that adversaries who compromised hosts can induce the controller to create a data-plane forwarding loop by only sending crafted packets in a specific order. This loop makes LoopGen a cost-effective DoS amplifier that can be combined with different DoS attacks. We conduct extensive experiments, evaluating LoopGen using diverse real-world topologies, different open-source controllers, and heterogeneous switches. The results show that LoopGen has non-trivial feasibility, significant amplification effect, low attack cost, and high stealthiness under existing defenses. Finally, we propose two countermeasures to mitigate this attack.

Tracegram: Framing Trace-Level Traffic Analysis with Temporally-Aware Multiple Instance Learning

Jian Qu, Yuchen Zhang, Jialong Zhang, and Jianfeng Li, MOE Key Laboratory for Intelligent Networks and Network Security, Xi’an Jiaotong University; Xiaobo Ma, MOE Key Laboratory for Intelligent Networks and Network Security and Shaanxi Province Key Laboratory of Computer Network, Xi’an Jiaotong University

Available Media

Modern network behaviors span multiple flows and evolve over time, making temporal and co-occurrence contexts across flows essential for reliable traffic analysis. This need is especially critical in the security domain, where attacks progress through reconnaissance, delivery, command and control, and lateral movement over extended intervals and across multiple flows. Existing packet-level or single-flow approaches fragment this context and limit performance on trace-level classification, detection, and attribution. We introduce the trace as the analysis unit and present Tracegram, which formulates trace-level analysis as Multiple Instance Learning. Tracegram combines per-flow encoders with a temporally aware aggregation module to reason across flows, preserve long-range dependencies, and produce key-flow attribution signals that support analyst verification and forensics. Our validation spans theory and practice. We theoretically justify the MIL-based decomposition for trace-level traffic analysis and conduct extensive experiments on four public datasets across multiple tasks, showing better or comparable performance to state-of-the-art methods. Finally, case studies on APT traces from the DAPT dataset show that Tracegram highlights flows aligned with attack phases, enabling targeted investigation.

The Ivory Tower Syndrome: Operators' Reflections on Academic BGP Security Solutions

Aleeza Suhel Inamdar, Saarland University; Tobias Fiebig, TU Wien; Mannat Kaur, Max-Planck-Institut für Informatik

Available Media

The Border Gateway Protocol (BGP) is the backbone of Internet routing, primarily used for exchanging network reachability information between autonomous systems on the Internet. BGP leverages a trust-based framework, where networks rely on each other's routing information implicitly. However, the initial design of BGP lacks in-built security mechanisms, thus making it vulnerable to various attacks and possibly network failures. Over the years, academics have introduced several solutions to enhance BGP security and address these vulnerabilities such as Resource Public Key Infrastructure (RPKI), Secure Border Gateway Protocol, and BGPsec. However, academic solutions may not always be readily applied by network operators. It is, therefore, important to understand the existing challenges in the transition from academic research to widespread implementation. We engaged with 12 network operators via online focus groups, focusing on their work experiences in implementing various BGP security mechanisms. All participants noted a gap between academic research and industry needs, indicating misaligned priorities. To bridge this gap, they suggested more collaboration between the two domains -- particularly through increased researcher involvement in the network operators' community to understand practical challenges. We propose ways to foster such collaboration and urge the research community to reflect on the purpose behind publishing.

ASRogue: Manipulating ASRank-Inferred AS Relationships

Yi Xu, Yihao Chen, Ke Xu, Qi Li, and Jianping Wu, Tsinghua University

Available Media

Accurate knowledge of Autonomous System (AS) business relationships is critical for many Internet security and networking tasks. Since these relationships are not publicly visible, operators and researchers rely on algorithms that infer them from empirical routing data. Yet, despite decades of widespread deployment, the security of AS relationship inference against adversarial manipulation remains largely underexplored. To close this gap, we conduct a systematic security analysis of ASRank, a foundational algorithm that underpins many established methods and widely used datasets. We show that ASRank's output is highly sensitive to small, targeted perturbations of AS triplets extracted from BGP paths. Carefully chosen triplet changes can alter inference order and trigger cascading misclassifications. Building on this insight, we present ASRogue, a manipulation attack against ASRank-based inference. ASRogue adaptively crafts AS triplets to steer inference order and outcomes, scales triplet injection via prefix splitting and AS-path poisoning, and embeds forged triplets into policy-compliant BGP announcements that can evade common routing defenses. Extensive evaluation shows that when the attacker controls at least two ASes, ASRogue achieves an overall 96.7% manipulation success rate, exceeding 99% for small and medium-sized providers, and remains effective under constrained attack overhead and temporal variation. Finally, controlled experiments on the PEERING testbed validate the feasibility of launching ASRogue attacks in the real-world Internet.

Understanding Implicit Trust Errors in Core Carrier Networks through Multi-Agent Flaw Discovery and Analysis

Ziyu Lin, Ziting Wang, Xinfeng Li, Wei Dong, and XiaoFeng Wang, Nanyang Technological University

Available Media

Core network (CN) is at the center of a cellular system and was historically protected by physical isolation, but the shift to cloud-native deployments breaks this protection and introduces new attack interfaces. Indeed, from reported GitHub issues, we found that many security flaws within CNs can be attributed to blind trust underlying interactions across different CN components. Under such trust assumptions, one component may fail to properly verify the syntactic and semantic content of messages received from other internal components, and may also grant resources to requesters without validation of resource availability. These weaknesses, once coupled with the risks of exposing internal interfaces to unauthorized external parties, as observed in our research, will lead to serious consequences, including session hijacking and denial of services. In our research, we call these newly discovered weaknesses implicit trust errors or iTrue.

To detect iTrues and understand their security impacts, we designed and implemented a multi-agent framework, dubbed iFinder. iFinder automatically analyzes and summarizes known security flaws, and uses them as ``seeds'' to detect related vulnerabilities from the CN simulators. To suppress hallucinations produced by large language models (LLMs), not only have we developed a Chain of Thought (CoT) reasoning to guide agents, but we also build an innovative strategy that cross-checks both 3GPP specifications and CN code to capture existing protection missed by the agents. Further, we developed a technique that uses LLMs to generate proof-of-concept (PoC) exploits for potential iTrues and iteratively refine the PoCs by automatically executing them against CN simulators and analyzing results. Running iFinder on 7 prominent simulators, we discovered 84 previously unknown vulnerabilities. Among them 70 have already been confirmed with 40 CVE assignments already awarded. Our findings highlight the pervasiveness of iTrue risks across the cellular core networks and the urgent needs for elevated protection within the original trust domains. We plan to make the service of iFinder publicly available to help enhance the security qualities of cellular core networks.

DNS Cache Poisoning Like it’s 2006

Omer Ben-Simhon and Amit Klein, Hebrew University of Jerusalem

Available Media

The Domain Name System (DNS) underpins virtually all Internet services, making the integrity of DNS resolution critical to security and availability. We present a comprehensive study of a novel class of DNS cache poisoning attacks against BIND9,the most widely deployed open-source DNS resolver. Our attack focuses on two keycapabilities that set it apart from most prior work: (1) reliably predicting both critical challenge parameters– the UDP source port and TXID– whereas most existing attacks target only one; and (2) performing this prediction entirely from the client side, without attacker-operated authoritative servers for attacker domains, which to our knowledge is a first. We achieve this by exploiting weaknesses in BIND's pseudo-random number generation, enabling highly reliable prediction even under realistic network conditions. In addition to the client-side-only techniques, we also develop server-side techniques which are needed in order to attack the older 9.18 branch of BIND 9. We evaluate our attacks and demonstrate practical success rates across multiple BIND 9 release branches and configurations. All vulnerabilities were responsibly disclosed to the Internet Systems Consortium (ISC) and the FreeBSD Project, leading to two patches and CVEs and acknowledgments.

iROV: Breaking the Silence of RPKI with Interactive Validation

Younsoo Kim and Seungjin Baek, KAIST; Weitong Li and Tijay Chung, Virginia Tech; Min Suk Kang, KAIST

Available Media

While Route Origin Validation (ROV) effectively mitigates prefix hijacking, it drops BGP updates that are inconsistent with currently published ROAs without consulting the prefix owner. This silent-drop behavior creates a serious agility gap: when a resource owner must urgently redirect traffic during DDoS mitigation, the temporary route may deviate from the currently published ROAs and be dropped before a ROA update propagates. We propose iROV (interactive ROV), a lightweight extension to ROV, to address the agility gap by transforming validation from a static local decision into an interactive, owner-involved process. iROV enables routers to query prefix owners for "Just-in-Time" authorization of invalid routes. This interactivity also allows prefix owners to receive real-time visibility feedback on ROV-invalidation events, which helps them detect remote prefix-hijacking attempts or quickly discover their own benign misconfigurations. To mitigate availability threats introduced by this interactivity, iROV employs validity stapling to suppress redundant queries, and probabilistic querying to bound owner-side query load. Large-scale simulations and a prototype implementation demonstrate that iROV restores reachability for urgent route changes from 23% to up to nearly 90%, and delivers rapid feedback on anomalies with granular impact estimation, while incurring negligible overhead.

Side-Channel Attacks on Open vSwitch

Daewoo Kim and Sihang Liu, University of Waterloo

Available Media

Virtualization is widely adopted in cloud systems to manage resource sharing among users. A virtualized environment usually deploys a virtual switch within the host system to enable virtual machines to communicate with each other and with the physical network. The Open vSwitch (OVS) is one of the most popular software-based virtual switches. It maintains a cache hierarchy to accelerate packet forwarding from the host to virtual machines. We characterize the caching system inside OVS from a security perspective and identify three attack primitives. Based on the attack primitives, we present three remote attacks via OVS, breaking the isolation in virtualized environments. First, we identify remote covert channels using different caches. Second, we present a novel header recovery attack that leaks a remote user's packet header fields, breaking the confidentiality guarantees from the system. Third, we demonstrate a remote packet rate monitoring attack that recovers the packet rate of a remote victim. To defend against these attacks, we also discuss potential mitigations.

DaLens: Charting DNS Self-Amplification Threats at Large

Liwen Xu and Zechao Cai, ETH Zurich; Huayi Duan, HKUST(GZ); Adrian Perrig, ETH Zurich

Available Media

The emerging self-amplification attacks (SAAs) pose serious denial-of-service (DoS) risks to the Domain Name System (DNS). They can substantially amplify the interactions between recursive and authoritative servers, depleting resources at disproportionally small costs. Assessing the impact of such attacks on the global name resolution infrastructure is crucial for DNS operators to effectively triage threats and deploy defenses, yet this remains an uncharted and daunting territory.

We have conducted the first large-scale measurement study of SAAs, leveraging a versatile framework, DaLens, which we designed and developed. The work consists of untangling the intricate ∈fra infrastructure to identify effective amplifiers and quantifying their amplification capabilities in a modular, scalable, and sound manner. Out of 307K persistent public resolvers, we find 29K unique resolver clusters that can be exploited in parallel for SAAs, and a significant number of them can still produce large amplification effects even though these vulnerabilities had already been disclosed in prior work.

12:30 pm–2:00 pm

Lunch (on your own)

2:00 pm–3:30 pm

Track 1

Applications of Cryptography 7

Session Chair: Eyal Ronen, Tel Aviv University

Grand Ballroom VII-X

Secure Distributed RSA Modulus Generation Revisited: A Faster and More Communication-Efficient Construction

Shaojing Zhang, College of Computer Science and Technology, Qingdao University; Chengliang Tian, College of Computer Science and Technology, Qingdao University, and State Key Laboratory of Cryptography and Digital Economy Security, Shandong University; Ruixue Wang and Hequn Xian, College of Computer Science and Technology, Qingdao University; Guangwu Xu, State Key Laboratory of Cryptography and Digital Economy Security and School of Cyber Science and Technology, Shandong University

Available Media

Securely generating an RSA modulus N = p q N=pq in a distributed n n-party setting while keeping its factors private is a core challenge in threshold cryptography. This paper revisits this problem and improves the state‑of‑the‑art distributed Miller–Rabin primality test by replacing its core component—the divisibility test—with a novel zero‑testing protocol.

Our key innovation is an elegant distributed design of Montgomery reduction, which transforms checking δmodf=0 (given shares of δ and f) into testing whether a product equals zero. This approach enables efficient distributed evaluation and offers: (1) Perfect correctness. Unlike prior approaches that are probabilistic or require impractically large shares, our zero test decides δmodf=0 deterministically. (2) Improved efficiency. Our zero test runs in only 7 rounds and uses 5 secure multiplications, improving over previous divisibility tests that require either 3+⌈log2n⌉ rounds with n+3 multiplications or 8 rounds with 5n−1 multiplications. Extensive experiments under honest-majority settings show increasing speedups for n ≥ 9, with the efficiency advantage growing as the number of parties increases. Moreover, our semi-honest protocol supports both honest- and dishonest-majority settings via appropriate secure multiplication primitives, and can be upgraded to malicious security with standard compilers.

Efficient Threshold ML-DSA

Sofía Celi, Brave Research and University of Bristol; Rafael del Pino and Thomas Espitau, PQShield; Guilhem Niot, PQShield and Univ Rennes, CNRS, IRISA; Thomas Prest, PQShield

Available Media

Threshold signature schemes allow a group of users to jointly generate a digital signature, providing resilience against faults and enhancing decentralization. With the advent of post-quantum cryptography, lattice-based threshold signatures have gained attention as viable alternatives. Nevertheless, existing constructions frequently encounter challenges related to scalability, robustness, or compatibility with standardized schemes, particularly with the NIST-selected and standardised Module-Lattice-based Digital Signature Algorithm (ML-DSA) algorithm.

In this work, we present the first threshold signature scheme that is fully compatible with ML-DSA, supporting secure and efficient signing among up to six parties. Our construction leverages advanced short secret sharing techniques and integrates optimized rejection sampling to achieve a favourable balance between communication efficiency and correctness in distributed environments. We implement our construction in Go and evaluate its performance across local, LAN, and WAN network settings. Our benchmarks demonstrate that our threshold ML-DSA scheme is not only practically deployable but also well-suited for real-world applications, including multi-device cryptocurrency wallets, threshold-based TLS authentication, and for Tor's directory authorities.

Concretely Efficient Blind Signatures Based on VOLE-in-the-Head Proofs and the MAYO Trapdoor

Carsten Baum and Marvin Beckmann, Denmark Technical University; Ward Beullens, IBM, Zürich; Shibam Mukherjee, Graz University of Technology and Know Center; Christian Rechberger, Graz University of Technology and TACEO

Available Media

Blind signatures (Chaum, CRYPTO 82) are important building blocks in many privacy-preserving applications, such as anonymous credentials or e-cash schemes. Recent years saw a strong interest in building Blind signatures from post-quantum assumptions, primarily from lattices. While performance has improved, no construction has reached practical efficiency in terms of computation and communication. The state of the art requires at least 20 KB size of communication for each showing of a lattice-based Blind signature to a verifier, and more than 100 ms in prover time.

In this work, we propose an alternative direction with a plausibly post-quantum Blind signature scheme called PoMFRIT. It builds on top of the VOLE-in-the-head Zero-Knowledge proof system (Baum et al. CRYPTO 2023), which we combine with the MAYO digital signature scheme (Beullens, SAC 2021). We implement multiple versions of PoMFRIT to demonstrate security and performance trade-offs, and provide detailed benchmarks of our constructions. Signature issuance requires (0.45) KB communication for Blind signatures of size (6.7) KB. Showing a Blind signature can be done in <76 ms even for a conservative construction with 128 bit security. As a building block for our Blind signature scheme, we implement the first VOLE-in-the-head proof for hash functions in the SHA-3 family, which we consider of independent interest.

Distributed Vector Commitments and Their Applications

Rui Gao and Huaqun Wang, Nanjing University of Posts and Telecommunications, The State Key Laboratory of Tibetan Intelligence; Zhiguo Wan, Hangzhou Normal University; Yuncong Hu, Shanghai Jiao Tong University

Available Media

Vector commitment (VC) schemes enable a prover to commit to a vector and later open any position with a short proof. However, existing VC schemes are designed for centralized settings, and cannot work in decentralized systems, where the input vector is distributed across multiple machines. Similarly, traditional VC schemes cannot leverage distributed parallel computation across multiple machines for acceleration.

To tackle this issue, we introduce a new notion—distributed VC (DVC), which allows multiple machines, each holding only a subvector of the input vector, to collectively commit to the entire vector and generate position proofs in a distributed manner. To the best of our knowledge, there is no prior work on DVCs and no existing work can trivially derive an efficient DVC scheme. The key challenge is that both commitments and proofs depend on the entire vector, while no single machine holds the complete vector in distributed settings.

We propose the first DVC scheme, HLE-DVC, which leverages M machines to process the distributed vector v of length N inparallel, with each machine holding a subvector of length N/M . HLE-DVC achieves compact proof size-O(logM) and allows each machine to generate all its position proofs in a single communication round,with communication cost O(logM) and computation cost O(NlogN/M). Moreover, HLE-DVC supports batch proving, proof aggregation, and efficient updates. W econduct the experiments and open-source the code. Using 256 machines to generate all proofs for a committed vector of length 230 takes 17,515 seconds. This achieves a 256× parallel speedup over HLE-DVC on a single machine, and is 142× faster than Hyperproofs (a famous single machine VC scheme). The communication cost per machine is 0.768 KB.

Heli: Heavy-Light Private Aggregation

Ryan Lehmkuhl and Henry Corrigan-Gibbs, MIT; Emma Dauterman, Stanford; David J. Wu, University of Texas at Austin

Available Media

This paper presents Heli, a system that lets a pair of servers collect aggregate statistics about private client-held data, without learning anything more about any individual client's data. Like prior systems, Heli protects client privacy against a malicious server, protects correctness against misbehaving clients, and supports common statistical functions: average, variance, and more. Heli's innovation is that only one of the servers (the "heavy server") needs to do per-run work proportional to the number of clients; the other server (the "light server") does work independent of the number of clients, after a one-time setup phase. As a result, a computationally limited party, such as a low-budget non-profit, could potentially serve as the second server for a Heli deployment with millions of clients.

Heli relies on a new cryptographic primitive, aggregation-only encryption, that allows computing certain restricted functions on many clients' encrypted data. In a deployment with ten million clients, in which the servers privately compute the sum of 32 client-held 1-bit integers, Heli's heavy server does 240,000 core-s of work and the light server does 7 core-ms of work. Compared with prior work, the heavy server does 38× more computation, but the light server does 120, 000× less.

Differential Trust: Dynamic Multi-Authority Anonymous Credentials with Epoch-Weighted Updates

Chen Li, Tianjin University; Jianting Ning, Zhejiang Sci-Tech University; Xiulong Liu, Tianjin University; Yulin Liu, Wuhan University

Available Media

Anonymous credentials (ACs) are fundamental to privacy-preserving authentication, allowing users to prove possession of attributes without revealing their identities. State-of-the-art ACs distribute credential issuance across multiple authorities, typically employing techniques such as Shamir's secret sharing or aggregate signatures. While this approach enhances system robustness and eliminates single point of failure, it treats all authorities equally in the credential issuance phase. This uniform treatment disregards the varying levels of trustworthiness or stake held by different authorities. Such limitation has become particularly problematic in modern decentralized systems like Proof-of-Stake networks, where the inherent trust differentiation among nodes cannot be leveraged in the credential issuance process.

To address this limitation, we propose the notion of Multi-Authority Anonymous Credentials with Epoch-Based Weights (MA-ACEW), the first Multi-Authority Anonymous Credential (MA-AC) model that considers authorities' weight distribution in credential issuance. Crucially, MA-ACEW enables efficient credential updates when authority weight distributions change across epochs. The core of MA-ACEW is our novel Epoch-Bound Pointcheval-Sanders Signature (EB-PS) primitive, which binds signatures to specific time epochs. This temporal binding enables both weight-based credential issuance within epochs and efficient non-interactive credential updates across epochs. We formalize the EUF-eCMA unforgeability requirement for EB-PS and prove our construction satisfies it under a novel STB-GPS assumption. We then prove that our MA-ACEW construction achieves unforgeability, anonymity, and blindness. Finally, we present benchmarks demonstrating the efficiency of EB-PS and MA-ACEW. Remarkably, presenting a credential aggregated from 128 partial ones takes only 10.68 ms on average.

Quorus: Efficient, Scalable Threshold ML-DSA Signatures from MPC

Alexander Bienstock, Leo de Castro, Daniel Escudero, Antigoni Polychroniadou, and Akira Takahashi, J.P. Morgan AlgoCRYPT CoE and J.P. Morgan AI Research

Available Media

A threshold signature protocol divides a secret signing key among multiple parties, enabling any subset above a threshold to jointly create a signature. While post-quantum (PQ) threshold signatures are being studied, especially following NIST's call for threshold schemes, most solutions focus on specially designed, threshold-friendly signature schemes. However, real-world applications like distributed certificate authorities and digital currencies require signatures verifiable under existing standardized procedures. With NIST's standardization of PQ signatures and ongoing industry deployment, designing an efficient threshold scheme compatible with NIST-standardized verification remains a critical challenge.

In this work, we present the first efficient and scalable solution for multi-party generation of the module-lattice digital signature algorithm (ML-DSA), one of NIST's PQ signature standards. Our contributions are two-fold. First, we present a variant of the ML-DSA signing algorithm that is amenable to efficient multi-party computation (MPC) and prove that this variant achieves the same security as the original ML-DSA scheme. Second, we present several efficient & scalable MPC protocols to instantiate the threshold signing functionality. Our protocols can produce threshold signatures with as little as 150 KB (per party) of online communication per rejection-sampling round. In addition, we instantiate our protocols in the honest-majority setting, which allows us to avoid any additional public key assumptions.

Our signatures verify under the same ML-DSA implementation for all security levels, with signature and verification key sizes matching ML-DSA; previous lattice-based threshold schemes could not match both of these sizes. Our solution provides the first method for producing threshold post-quantum signatures compatible with NIST-standardized verification, scalable to any number of parties, without new assumptions.

Anonymous Tokens with Designated-Reader Metadata Bit

Aisha Tu, Wuhan University; Meng Jia, The Hong Kong Polytechnic University; Kun He, Jing Chen, and Ruiying Du, Wuhan University

Available Media

Anonymous tokens with private metadata bit convey hidden signals to verifiers when presented by the user and are under discussion in standardization. Existing solutions only allow the token issuer to read the signals, which places a heavy burden on the issuer and makes it challenging to support issuer-hiding because verifiers have to contact the issuer. In this paper, we propose an anonymous token scheme with designated-reader metadata bit, allowing the user to specify an issuer-accepted verifier to read the signal from the token directly. We also extend our scheme to support reader-hiding, which conceals the user's intended verifier from the issuer and other verifiers, and issuer-hiding, which prevents exposure of the token issuer from verifiers. We prove the security of our constructions and report their performance.

Streaming Function Secret Sharing and Its Applications

Xiangfu Song, Nanyang Technological University; Jianli Bai, Singapore Management University; Ye Dong, National University of Singapore; Yijian Liu, Yu Zhang, and Xianhui Lu, Institute of Information Engineering, Chinese Academy of Sciences and School of Cyber Security, University of Chinese Academy of Sciences; Tianwei Zhang, Nanyang Technological University

Available Media

Collecting statistics from users of software and online services is crucial to improve service quality, yet obtaining such insights while preserving individual privacy remains a challenge. Function secret sharing (FSS) is a promising tool for this problem. However, FSS-based solutions still face several challenges for streaming analytics, where messages are continuously sent, and secure computation tasks are repeatedly performed over incoming messages. We introduce a new cryptographic primitive called streaming function secret sharing (SFSS), a new variant of FSS that is particularly suitable for secure computation over streaming messages. We formalize SFSS and propose concrete constructions, including SFSS for point functions, predicate functions, and feasibility results for generic functions. SFSS powers several promising applications in a simple and modular fashion, including conditional transciphering, policy-hiding aggregation, and attribute-hiding aggregation. In particular, our SFSS formalization and constructions identify security flaws and efficiency bottlenecks in existing solutions, and SFSS-powered solutions achieve the expected security goal with asymptotically and concretely better efficiency and/or enhanced functionality.

Track 2

Privacy of ML 2

Session Chair: Xiaoyong (Brian) Yuan, Clemson University

Grand Ballroom VI

Connect the Dots: Knowledge Graph–Guided Crawler Attack on Retrieval-Augmented Generation Systems

Mengyu Yao, MOEKey Lab of HCST (PKU), School of Computer Science, Peking University; Ziqi Zhang and Ning Luo, University of Illinois Urbana-Champaign; Shaofei Li, Yifeng Cai, Xiangqun Chen, Yao Guo, and Ding Li, MOEKey Lab of HCST (PKU), School of Computer Science, Peking University

Available Media

Stealing attacks pose a persistent threat to the intellectual property of deployed machine-learning systems. Retrieval-augmented generation (RAG) intensifies this risk by extending the attack surface beyond model weights to knowledge base that often contains IP-bearing assets such as proprietary runbooks, curated domain collections, or licensed documents. Recent work shows that multi-turn questioning can gradually steal corpus content from RAG systems, yet existing attacks are largely heuristic and often plateau early. We address this gap by formulating {RAG knowledge-base stealing} as an adaptive stochastic coverage problem (ASCP), where each query is a stochastic action and the attacker's goal is to maximize the conditional expected marginal gain (CMG) in corpus coverage under a query budget. Bridging ASCP to real-world black-box RAG knowledge-base stealing raises three challenges: CMG is unobservable, the natural-language action space is intractably large, and feasibility constraints require stealthy queries that remain effective under diverse architectures. We introduce RAGCrawler, a knowledge graph-guided attacker that maintains a global attacker-side state to estimate coverage gains, schedule high-value semantic anchors, and generate non-redundant natural queries. Across four corpora and four generators with BGE retriever, RAGCrawler achieves 66.8% average coverage (up to 84.4%) within 1,000 queries, improving coverage by 44.90% relative to the strongest baseline. It also reduces the queries needed to reach 70% coverage by at least $4.03x on average and enables surrogate reconstruction with answer similarity up to 0.699. Moreover, the attack remains effective under retriever switching and newer RAG techniques such as query rewriting and multi-query retrieval, while being difficult for existing defenses to block. These results highlight urgent needs to protect RAG knowledge assets.

Network-Level Prompt and Trait Leakage in Local Research Agents

Hyejun Jeong, Mohammadreza Teymoorianfard, Abhinav Kumar, Amir Houmansadr, and Eugene Bagdasarian, University of Massachusetts Amherst

Available Media

We show that Web and Research Agents (WRAs)—language-model-based systems that investigate complex topics on the Internet—are vulnerable to inference attacks by passive network observers. Deployment of WRAs locally by organizations and individuals for privacy, legal, or financial purposes exposes them to DNS resolvers, malicious ISPs, VPNs, web proxies, and corporate or government firewalls. However, unlike sporadic and scarce web browsing by humans, WRAs visit 70-140 domains per each request with a distinct timing pattern creating unique privacy risks.

Specifically, we demonstrate a novel prompt and user trait leakage attack against WRAs that only leverages their network-level metadata (i.e., visited IP addresses and their timings). We start by building a new dataset of WRA traces based on real user search queries and queries generated by synthetic personas. We define a behavioral metric (called OBELS) to comprehensively assess similarity between original and inferred prompts, showing that our attack recovers over 73% of the functional and domain knowledge of user prompts. Extending to a multi-session setting, we recover up to 19 of 32 latent traits with high accuracy. Our attack remains effective under partial observability and noisy conditions. Finally, we discuss mitigation strategies that constrain domain diversity or obfuscate traces, showing negligible utility impact while reducing attack effectiveness by an average of 29%.

VOID: Defeating Unauthorized Mimicry in Latent Diffusion Models

Chunlin Qiu, Ang Li, Tianxiao Huang, Ruilin Gan, Yunjie Ge, and Shenyi Zhang, Wuhan University; Huayi Duan, The Hong Kong University of Science and Technology (Guangzhou); Lingchen Zhao, Wuhan University; Chao Shen, Xi'an Jiaotong University; Qian Wang, Wuhan University

Available Media

While Latent Diffusion Models (LDMs) have revolutionized visual synthesis, they are increasingly exploited for unauthorized mimicry of individuals. Existing defenses inject deceptive perturbations to steer the generated images toward irrelevant targets. However, this approach hinges on an ungrounded assumption: subtle perturbations can maintain their deceptive efficacy throughout an LDM's extensive generation process. In reality, the model's innate restoration mechanism will remove such perturbations and cause individual identities to re-emerge in the images generated.

We propose VOID, a defense framework that overcomes this conundrum by manipulating an LDM's intrinsic stochasticity. VOID perturbs the diffusion pipeline in two novel ways: 1) amplifying the latent encoding errors to shatter an image's semantic structure, and 2) counteracting the target guidance signals to suppress the model's restoration capabilities. This results in a semantic corruption that thwarts any unauthorized mimicry. Notably, the security gain does not come at the price of visual utility, as VOID simultaneously manages to confine perturbations to human-imperceptible regions of protected images. Our comprehensive evaluation of 24 state-of-the-art defenses against 10 mimicry attacks on 5 datasets demonstrates VOID's unprecedented protection power: it increases the average Fréchet Inception Distance (FID) from 113 to 365, a 223% improvement over the strongest defense to date.

From Easy to Hard++: Promoting Differentially Private Image Synthesis Through Spatial-Frequency Curriculum

Chen GONG and Kecen Li, University of Virginia; Zinan Lin, Microsoft Research; Tianhao Wang, University of Virginia

Available Media

Differentially private (DP) synthetic images serve as a critical tool for reducing privacy concerns by mimicking the statistical properties of sensitive data while ensuring privacy guarantees. To improve the quality of synthetic images, most studies have focused on improving the core optimization techniques (e.g., DP-SGD). Recently, we have witnessed a paradigm shift that takes these techniques off the shelf and studies how to use them together to achieve the best results. One notable work is DP-FETA, which proposes using 'central images' for 'warming up' the DP training and then using traditional DP-SGD.

Inspired by DP-FETA, we are curious whether there are other such tools we can use together with DP-SGD. We first observe that using 'central images' only works for datasets where there are many samples that look similar. To handle scenarios where images could vary significantly, we propose FETA-Pro, which introduces frequency features as 'training shortcuts.' The complexity of frequency features lies between that of spatial features (captured by 'central images') and full images, allowing for a finer-grained curriculum for DP training. To incorporate these two types of shortcuts together, one challenge is to handle the training discrepancy between spatial and frequency features. To address it, we leverage the pipeline generation property of generative models (instead of having one model trained with multiple features/objectives, we can have multiple models working on different features, then feed the generated results from one model into another) and use a more flexible design. Specifically, FETA-Pro introduces an auxiliary generator to produce images aligned with noisy frequency features. Then, another model is trained with these images, together with spatial features and DP-SGD. Evaluated across five sensitive image datasets, FETA-Pro shows an average of 25.7% higher fidelity and 4.1% greater utility than the best-performing baseline, under a privacy budget ε = 1.

CompLeak: Deep Learning Model Compression Exacerbates Privacy Leakage

Na Li, School of Cyber Science and Engineering, Nanjing University of Science and Technology, China; Yansong Gao, School of Cyber Science and Engineering, Southeast University, China; Hongsheng Hu, School of Computer Science, Shanghai Jiao Tong University, China; Boyu Kuang, School of Cyber Science and Engineering, Nanjing University of Science and Technology, China; Anmin Fu, School of Cyber Science and Engineering, Nanjing University of Science and Technology, China; and School of Computer Science and Engineering, Nanjing University of Science and Technology, China

Distinguished Artifact Award Winner

Available Media

Model compression is crucial for minimizing memory storage and accelerating inference in deep learning (DL) models. Users can access different compressed model versions according to their resources and budget. However, while existing compression operations primarily focus on optimizing the trade-off between resource efficiency and model performance, the privacy risks introduced by compression remain overlooked and insufficiently understood.

In this work that focuses on typical classification tasks, through the lens of membership inference attack (MIA), we propose CompLeak, the first privacy risk evaluation framework examining three widely used compression configurations that are pruning, quantization, and weight clustering all supported by the commercial model compression framework of Google's TensorFlow-Lite (TF-Lite), and first two supported by Facebook's PyTorch Mobile and the open-source toolkit of Microsoft NNI. CompLeak has three variants, given access to the available number of compressed models and/or the original model. CompLeakNR starts by adopting existing MIA methods to attack each individual compressed model, and identifies that different compressed models influence members and non-members differently. When the original model and one compressed model are available, CompLeakSR leverages the compressed model as a reference to the original model and uncovers more privacy by combining meta information (e.g., confidence vector) from both models. When multiple compressed models are available with/without accessing the original model, CompLeakMR innovatively exploits privacy leakage info from multiple compressed versions to substantially signify the overall privacy leakage. We conduct extensive experiments on six diverse model architectures (from ResNet to BERT and GPT-2), and five image and textual benchmark datasets. Our experimental results show that CompLeakMR achieves the best MIA performance on all evaluation metrics, including TPR @ 0.1% FPR, proving that model compression exacerbates privacy leakage.

Cracking Federated Privacy: Initialization-Resilient Gradient Inversion with Fine-Grained Reconstruction

Kaiming Zhu, Jinsheng Yang, Siyang Guo, Huaqian Qin, Taiyu Wang, Junbo Wang, Yuhong Nan, and Zibin Zheng, Sun Yat-sen University, P.R. China

Available Media

Federated Learning (FL) remains vulnerable to Gradient Inversion Attacks (GIA), where shared gradients can reveal clients' private data. Existing attacks struggle under early-stage initialization variations and often produce coarse reconstructions. In this paper, we identify sparsity changes in shared gradients as the primary source of this sensitivity and propose an initialization-resilient GIA with a coarse-to-fine design, achieving fine-grained recovery. The coarse stage aligns gradient directions and constrains non-zero entries to mitigate sparsity changes, while the fine stage refines magnitude alignment by a hybrid metric combining Cosine distance with a deformed Manhattan term. Extensive experiments against five baselines show up to 200% PSNR gain (25.4 → 47.7 dB) under sensitive initializations on CIFAR-10/100, with consistently delivering fine-grained recovery across four datasets and the entire FL lifecycle. Our method maintains competitive performance with SOTA baselines across batch sizes and local steps and reveals persistent leakage on several popular models and insufficient defenses, underscoring the urgent need for stronger privacy-preserving mechanisms.

DP-SAPF: Saliency-Aware Parameter Fine-tuning of Public Models for Differentially Private Image Synthesis

Chen Gong, University of Virginia; Kecen Li, NUS; Zinan Lin, Microsoft Research; Tianhao Wang, University of Virginia

Available Media

Differentially private (DP) image synthesis generates images that preserve the statistical characteristics of a sensitive dataset, enabling sensitive data analysis and usage while providing rigorous guarantees of privacy leakage. Existing methods fine-tune public models using DP Stochastic Gradient Descent (DP-SGD) on sensitive images to generate synthetic images. But full fine-tuning public models on sensitive images is computationally expensive, because current public models typically contain a large number of parameters. Recent work proposes heuristically using Low-Rank Adaptation (LoRA) on all attention-layer parameters of public models to reduce the number of trainable parameters. However, we argue that exhaustive LoRA coverage across all attention-layer parameters is suboptimal in a DP setting, as it leads to noise accumulation and collapse during private training.

To address this issue, we propose DP-SAPF, which uses a saliency-aware strategy to identify specific target parameters for LoRA training under DP. DP-SAPF is inspired by the fact that larger gradients signify higher saliency, indicating that these parameters are most critical for the DP learning. Specifically, we feed the sensitive images into public models, compute gradients, and add noise to the gradients to satisfy DP. Then, DP-SAPF identifies the most salient parameters, those exhibiting high gradient magnitudes on sensitive images, for DP fine-tuning. Experiments on four sensitive image datasets show that DP-SAPF improves the utility and fidelity of synthetic images while requiring fewer computational resources than fine-tuning methods without parameter selection.

CTA: Clip-then-Aggregate, a Differentially Private Learning Algorithm for Graph Data

Jianxin Wei, Mohamed bin Zayed University of Artificial Intelligence

Available Media

Graph learning has become a fundamental tool for modeling relational data in applications such as social networks and recommender systems, yet it often involves individuals' private features or interactions. Differential privacy (DP) provides a rigorous framework for protecting sensitive data, but its application to graph learning is challenging due to complex inter-node dependencies induced by message aggregation. Despite this broad influence, the per-neighbor impact of a node is typically small. Nevertheless, most existing private graph learning methods are built upon the DP-SGD framework, whose reliance on global gradient clipping prevents the precise capture of these influences and overestimates the true gradient sensitivity.

To address these challenges, we propose CTA, a differentially private graph learning algorithm that enables end-to-end training while perturbing only model gradients with tightly characterized sensitivity. CTA adopts a clip-then-aggregate design in both the forward and backward passes of model training. In the forward pass, CTA clips node embeddings before aggregation. Crucially, in the backward pass, instead of clipping each gradient as a whole, CTA separately clips and bounds two gradient components and then aggregates them to form the final model gradients. This design allows CTA to capture fine-grained gradient variations induced by graph aggregation and derive tighter gradient sensitivity bounds. Extensive experiments on real-world datasets demonstrate that CTA consistently outperforms existing private graph learning methods across a wide range of settings.

MASLeak: Investigating and Exposing Intellectual Property Leakage Vulnerabilities in Multi-Agent Systems

Liwen Wang, The Hong Kong University of Science and Technology; Wenxuan Wang, Renmin University of China; Shuai Wang, Zongjie Li, Zhenlan Ji, and Zongyi Lyu, The Hong Kong University of Science and Technology; Daoyuan Wu, Lingnan University; Shing-Chi Cheung, The Hong Kong University of Science and Technology

Available Media

The rapid advancement of Large Language Models (LLMs) has led to the emergence of Multi-Agent Systems (MAS) to perform complex tasks through collaboration. However, the intricate nature of MAS, including their architecture, agent interactions, and complex internal communication processing, raises significant concerns regarding intellectual property (IP) protection. In this paper, we introduce MASLEAK, the first framework for systematically extracting IP from MAS in a practical black-box setting. We assume a realistic adversary who can only submit queries to the system's public API and observe the final output, without any prior knowledge of the internal architecture and the backend LLM information. Inspired by how computer worms propagate and infect vulnerable network hosts, MASLEAK carefully crafts adversarial query q to elicit, propagate, and retain responses from each MAS agent that reveal a full set of proprietary components, including the number of agents, topology, system prompts, task instructions, and tool usages. We construct the first synthetic dataset of 810 MAS applications and also evaluate MASLEAK against real-world MAS applications, including Coze and CrewAI. MASLEAK achieves high accuracy in extracting MAS IP, with an average attack success rate of 87% for system prompts and task instructions, and 92% for system architecture in most cases. We conclude by discussing the implications of our findings and the potential defenses.

Track 3

Human Aspects 4

Session Chair: Filipo Sharevski, DePaul University

Grand Ballroom V

From Assistance to Autonomy: An Empirical Study of AI Use in a Live Capture-the-Flag (CTF) Competition

Tingxuan Tang, Nicolas Janis, and Kalyn Asher Montague, William & Mary; Kevin Eykholt, Dhilung Kirat, Youngja Park, and Jiyong Jang, IBM Research; Adwait Nadkarni and Yue Xiao, William & Mary

Available Media

Capture-the-Flag (CTF) competitions are increasingly becoming a testbed for evaluating AI capabilities at solving security tasks, due to their controlled environments and objective success criteria. Existing evaluations have focused on how successful AI is at solving individual CTF challenges in isolation from human CTF players. As AI usage increases in both academic and industrial settings, it is equally likely that human CTF players may collaborate with AI agents to solve CTF challenges. This possibility exposes a key knowledge gap: how do human players perceive AI CTF assistance; when assistance is provided, in what ways do they collaborate and is it effective with respect to human performance; how do humans assisted by AI compare to the performance of fully autonomous AI agents on the same set of challenges. We address this gap with the first empirical study of AI assistance in a live, onsite CTF. In a study with 41 participants (out of the total 95 that participated in the CTF), we qualitatively study (i) how participants' perception, trust, and expectations shift before versus after hands-on AI use, and (ii) how participants collaborate with an instrumented AI assistant. Moreover, we also (iii) benchmark four autonomous CTF agents on the same fresh challenge set to compare outcomes with human teams and analyze agent trajectories. We find that, for human players, AI literacy and domain knowledge are complementary competencies, and both have irreplaceable advantages. Proficient and efficient use of AI amplifies professional skills. Importantly, although advanced autonomous agents showed outstanding performance, human-in-the-loop is the winning paradigm where AI accelerates exploration while humans provide targeted guidance and verification. We conclude with implications for the future design of CTF competitions and for building effective human-in-the-loop AI systems for security.

Chameleon Channels: Measuring YouTube Accounts Repurposed for Deception and Profit

Alejandro Cuevas, Carnegie Mellon University; Manoel Horta Ribeiro, Princeton University; Nicolas Christin, Carnegie Mellon University

Available Media

Online content creators spend significant time and effort building their user base through a long, often arduous process that requires finding the right "niche" to cater to. So, what incentive is there for an established content creator known for cat memes to completely reinvent their page channel and start promoting cryptocurrency services or covering electoral news events? And, if they do, do their existing subscribers not notice?

We explore this problem of repurposed channels, whereby a channel changes its identity and contents. We first characterize a market for "second-hand" social media accounts, which recorded sales exceeding USD 1M during our 6-month observation period. Observing YouTube channels (re)sold over these 6 months, we find that a substantial number (53%) are used to disseminate policy-sensitive content, often without facing any penalty. Even more surprisingly, these channels seem to gain rather than lose subscribers.

We estimate the prevalence of channel repurposing "in the wild," using two snapshots of  1.4M YouTube accounts sampled from an ecologically valid proxy. In a 3-month period, we estimate that  0.25% channels were repurposed. Through a set of experiments, we confirm that these repurposed channels share several characteristics with sold channels—mainly the fact that they have a significantly high presence of policy-sensitive content. Across repurposed channels, we find channels similar to those used in influence operations, as well as channels used for financial scams. Repurposed channels have large audiences; across two observed samples, repurposed channels collectively held  193M and  44M subscribers. We reason that purchasing an existing audience and the credibility associated with an established account is advantageous to financially- and ideologically motivated adversaries. This phenomenon is not exclusive to YouTube and we posit that the market for cultivating organic audiences is set to grow, particularly if it remains unchallenged by mitigations, technical or otherwise.

Missing, Present and Conflicting: A Large Scale Analysis of IoT Update Information in the EU Market

Swaathi Vetrivel, Michel van Eeten, and Carlos H. Gañán, Delft University of Technology

Available Media

Security updates are essential for protecting IoT devices, yet consumers often lack reliable information about how long devices will be supported. We conduct the first large-scale study of update duration disclosures in the European market, analysing 34,187 product pages across local retailers, EU Amazon sites, and Temu. Disclosure varies sharply: Dutch retailers, subject to regulatory oversight, list update durations for up to 92% of devices, while Amazon provides such information for fewer than 1% and Temu for none. For smart TVs, where EU rules mandate disclosure, coverage is higher but still inconsistent. Stated update durations vary between one and eight years, with smart TVs generally receiving the longest support. Comparing stated support durations across retailers, manufacturers, and the EU's central product database, we find widespread contradictions, with retailers often understating support relative to manufacturers. These inconsistencies limit the effectiveness of transparency mandates and risk misleading consumers. Our findings show that regulation can improve visibility, but only robust enforcement and standardized disclosure mechanisms ensure accurate and trustworthy information.

Not All Those Who Share Are Lost: Analyzing 25 Years of Cybersecurity Artifact Sharing Practices Through Automated Discovery

Daan Vansteenhuyse, Arthur Bols, and Lieven Desmet, DistriNet, KU Leuven; Victor Le Pochat, Independent; Jo Van Bulck and Marton Bognar, DistriNet, KU Leuven

Available Media

Cybersecurity publications are often accompanied by artifacts (code and data) that support the paper's findings. In recent years, all top-tier cybersecurity conferences have adopted policies regarding artifacts and introduced artifact evaluation. However, large-scale trends in artifact availability and the impact of conference policies remain largely unexplored, and meta-science studies (often focusing on reproducibility) typically cover a small set of papers due to the extensive manual effort required.

In this paper, we enable larger-scale analyses by introducing ArtiFinder, an automated tool that accurately identifies artifact URLs in paper PDFs. Using ArtiFinder, we present the largest quantitative analysis to date of artifact availability and related effects in close to 9,000 papers published in the four leading cybersecurity conferences since 2000. Our longitudinal analysis reveals a steady increase in artifact sharing over time, with substantial variation across cybersecurity subfields. We find that recent policy changes are not consistently accompanied by increases in artifact sharing, but they coincide with clear shifts toward stable hosting services. We also replicate prior analyses on our dataset, examining trends in citation counts and repository popularity. Our findings highlight notable trends in artifact sharing and the guiding role of conference organizers, while our open-source tool and dataset enable future large-scale studies.

Digital Risks and Coping Practices among Roblox Game Creators

Qiurong Song, Rie Helene (Lindy) Hernandez, Xinning Gui, and Yubo Kou, The Pennsylvania State University

Available Media

As a growing part of the creator economy, game platforms like Roblox enable millions of users to design, publish, promote, and monetize games. Alongside these opportunities, however, creators on such platforms face significant safety, privacy, and security risks. While prior work has examined online risks for content creators on social media platforms, little is known about the risk landscape of game creators. To address this gap, we interviewed 20 Roblox creators to understand how they perceive, experience, and cope with digital risks. Our analysis revealed five categories of risk—platform, production, organizational, community, and technical—that potentially compromise Roblox game creators' emotional, physical, relational, and financial safety. We also identified coping strategies such as negotiating for fairer pay and seeking community support. We conclude with recommendations for strengthening protections for game creators.

Do They Get With the Program? Measuring Mitigation in a Solicited Vulnerability Notification Program

Yana Angelova, Carlos Gañán, Annebel Smit, Rolf van Wegberg, and Michel van Eeten, Delft University of Technology

Available Media

Attackers rapidly scan for newly-disclosed vulnerabilities across enterprise networks. In response, defenders have been running similar scans in order to notify the affected enterprise about their exposed attack surface as soon as possible. Such initiatives are becoming more institutionalized, e.g., under the EU's NIS2 directive and CISA's Ransomware Vulnerability Warning Pilot. Research on the effectiveness of vulnerability notifications has found disappointing results, where only a minor fraction of issues were fixed. This was blamed on problems in asset attribution, low trust in the sender, reachability issues, and lacking incentives of the recipient to act. A potential solution to all these problems would be a notification program where companies volunteer to sign up, register their assets, and ensure the right contact details. How much better could such a program perform? We provide the first empirical evaluation of a solicited notification program through a collaboration with a governmental Computer Security Incident Response Team (CSIRT) that offers security notifications to enterprises. We first conduct interviews with nearly half of all participating companies (n = 21) to understand why they signed up and how they act on the notifications. Next, we quantitatively study the remediation effectiveness of the program via survival analysis. We find that 27% of the security issues being resolved within one day of notification, 40% within one week, and 49% within one month. Over the entire three years of the program, the remediation rate is 75%. These findings already show higher remediation effectiveness compared to previous unsolicited experiments. This suggests that solicited notification programs can overcome challenges of reachability, trust, and motivation. We reflect on the limitations of these findings and their implications for the future.

Security in the Air: Understanding IoT Vendor Practices and the Economics of Over-the-Air Updates

Huancheng Hu and Christian Doerr, Hasso Plattner Institute, University of Potsdam

Available Media

Over-the-air (OTA) firmware updates are essential for maintaining IoT device security. However, vendors frequently reduce update frequency and discontinue support before devices reach end of life. Prior work has focused on protocol design and deployment measurements, but has not explained why vendors struggle to sustain long-term update support. Through interviews with 30 IoT vendors and a survey of over 100 practitioners, we examine OTA practices from the vendor perspective. We find that OTA failures stem from fragmented responsibility across multi-tier supply chains, economic constraints that determine support duration independent of security needs, and structural mismatches between component lifecycles and deployment periods. These organizational and economic factors create systematic barriers to security maintenance that technical solutions alone cannot address. Our findings reveal why current approaches to IoT security are insufficient and suggest policy interventions for sustainable update practices.

"Sticking their heads out above the parapets": Lived Experiences of Legal Risks in Research

Sunoo Park, New York University; Daniel R. Thomas, University of Strathclyde

Available Media

Overbroad computer crime, intellectual property, and other laws are well known to create legal risks for essential research. Notable examples include the US Computer Fraud and Abuse Act and the UK Computer Misuse Act. Because such laws fail to distinguish malicious hacking from good-faith testing and research, researchers face serious legal risks for public-interest research activity like identifying vulnerabilities or scraping data. Despite the research community's broad awareness of these risks, our understanding of their practical impacts is limited, as most of the community's knowledge comes from anecdotal evidence rather than systematic study.

We conduct the first qualitative study focused on researchers' lived experiences, to empirically document the impacts of legal risks and threats on research and researchers, and how researchers navigate legal risk situations. Our study engages two groups: researchers (NR = 36), who discuss 130 projects and incidents, and professionals that offer support to researchers navigating legal risks (NS = 8), who have supported thousands of researchers. We thus provide an unprecedented big-picture view of researchers' experiences with legal risks. We synthesise actionable strategies for researchers, and our findings provide evidence to support policy reform.

SoK: Beyond Burnout -- A Gap Analysis of Psychological Harm in Cybersecurity Work

Jessica Phelan and Ivan Flechais, University of Oxford

Available Media

Burnout among cybersecurity professionals drives workforce attrition, but its consequences extend beyond staffing shortages. Degraded judgment and human error, particularly during incident response, can be catastrophic. Yet the sources and forms of psychological harm across cybersecurity roles remain poorly understood. This Systematization of Knowledge (SoK) paper examines the literature on wellbeing and psychological harm in cybersecurity work. Using a scoping review across five databases, we review 101 papers and synthesize 57 covering wellbeing research on cybersecurity practitioners and four adjacent professions selected for structural similarity. We find that research on offensive cyber roles emphasizes harm driven by operational conditions, while the work itself is often experienced as meaningful. In contrast, research on defensive roles highlights stressors intrinsic to the work, including sustained vigilance, exposure to human malice, and invisible success metrics. Adjacent professions facing similar demands consistently report psychological harm extending beyond burnout. How such harm manifests in cybersecurity practitioners remains unclear, with existing studies measuring burnout, cognitive workload, fatigue, clinical distress, or wellbeing, but no studies to date assessing whether psychological harms beyond these constructs are present in this population. We identify cybersecurity-specific barriers to processing occupational stress, including confidentiality constraints, limited institutional support, cultures discouraging help-seeking, and temporal dynamics that prevent psychological closure. Our findings provide a structured foundation for research to explore cybersecurity interventions calibrated to the psychological harms the work actually produces.

Track 4

System Security 3

Session Chair: Adam Doupé, Arizona State University

Grand Ballroom I-IV

Sliding into the Flight Deck's DMs: Practical Message Attacks on CPDLC

Mehdi Ziazi, ETH Zurich; Khalid Aleem, Independent; Harshad Sathaye, ETH Zurich; Martin Strohmeier, Cyber-Defence Campus, armasuisse Science + Technology

Distinguished Paper Award Winner

Available Media

The Controller–Pilot Data Link Communications (CPDLC) system has become integral to modern air traffic management, particularly in high-density or oceanic airspace where voice communication is limited or unavailable. Designed to increase operational efficiency, CPDLC is an alternative to traditional VHF voice communication with standardized digital messages for altitude changes, heading adjustments, free-text messages, and frequency handovers. However, CPDLC does not implement encryption and relies primarily on protocol complexity and obscurity as a barrier to misuse.

In this work, we present a full-stack security analysis of CPDLC and showcase several vulnerabilities that allow hijacking ATC-Pilot link with rogue ground station attacks and large-scale denial of service attacks that are capable of disabling CPDLC services for all aircraft in radio range. As a proof-of-concept, we also introduce cpdlc-gs, a first SDR based full-stack CPDLC ground-station implementation capable of injecting uplink messages to issue fake CPDLC flight instructions and effective denial of service attacks.

Furthermore, to evaluate cpdlc-gs, together with air navigation service providers and avionics manufacturers, we develop a novel, fully-functional test environment with real, certifiable hardware from Universal Avionics. Through such a setup we conceptualize and validate several attacks and demonstrate that even isolated rogue stations can pose a substantial threat, especially when pilots are under high workload or in degraded communication scenarios. Overall, we argue that the heavy reliance and global adoption of CPDLC make it a high value target, and that the lagging aviation datalink security standard- ization process needs to be urgently addressed

PatchWeaver: Risk-Bounded Autonomous Vulnerability Remediation Under Change-Management Policies

Rui Li and Shuang Cao, Dawn Computing

Available Media

Security teams increasingly rely on automation to keep pace with vulnerability disclosure and patch deployment, yet real remediation workflows are constrained by change-management policies: maintenance windows, staged rollout rules, blast-radius limits, approval gates, and least-privilege requirements. Today, "autonomous" remediation is brittle. Scripted playbooks are safe but rigid, while LLM-driven agents are flexible but frequently violate operational policies (e.g., patching outside windows, taking too many replicas offline, or skipping mandatory validation), creating outages and compliance risk.

We present PatchWeaver, a remediation system that provides policy-bounded autonomy. PatchWeaver combines (i) an explicit, continuously refreshed graph of Kubernetes assets, dependencies, identities, vulnerabilities, approvals, and evidence; (ii) a typed policy interface (ChangeSpec) that encodes organization-specific change rules as executable predicates; and (iii) a rollout-based decision layer that scores candidate remediation plans by predicting both progress and policy-violation risk before any action is executed. The core insight is that myopic enforcement cannot avoid constraint traps--states where locally admissible actions lead to dead ends that force later violations. PatchWeaver detects traps via short-horizon simulation and replans when reality diverges from prediction.

We evaluate PatchWeaver on four Kubernetes remediation workloads under explicit governance constraints. Compared to a strong toolchain baseline (Gatekeeper + Argo Rollouts + Cosign), PatchWeaver reduces step-level policy violations by 52.5% (4.73% vs. 9.95%, p<0.01) and episode-level violations by 52.2% (5.82% vs. 12.18%). Compared to an LLM-agent baseline, PatchWeaver reduces step violations by 79.1% and worst-case (p99) episode violations by 3.1x (16.8% vs. 51.3%). In security stress tests across 10 attack categories, PatchWeaver achieves 0.33% aggregate miss rate with mean dwell time 0.84 s, separating 88.7% immediate blocks from 11.0% escalations, versus 28.5% miss rate and 96.4 s for the toolchain. Control-plane overhead remains practical: p99 admission latency is 31 ms at 32 policy predicates, and total resource usage is under 1 CPU core and 1 GB memory.

Trustworthy and Confidential SBOM Exchange

Eman Abu Ishgair and Chinenye Okafor, Purdue University; Marcela S. Melara, Intel Corporation; Santiago Torres-Arias, Purdue University

Available Media

Software Bills of Materials (SBOMs) have become a regulatory requirement for improving software supply chain security and trust by means of transparency regarding components that make up software artifacts. However, enterprise and regulated software vendors commonly wish to restrict who can view confidential software metadata recorded in their SBOMs due to intellectual property or security vulnerability information. To address this tension between transparency and confidentiality, we propose Petra, an SBOM exchange system that empowers software vendors to interoperably compose and distribute redacted SBOM data using selective encryption. Petra enables software consumers to search redacted SBOMs for answers to specific security questions without revealing information they are not authorized to access. Petra leverages a format-agnostic, tamper-evident SBOM representation to generate efficient and confidentiality-preserving integrity proofs, allowing interested parties to cryptographically audit and establish trust in redacted SBOMs. Exchanging redacted SBOMsin our Petra prototype requires less than 1 extra KB per SBOM, and SBOMdecryption accounts for at most 1% of the performance overhead during an SBOM query.

Zombie Cards Back Online: Reviving Expired Credit Cards for Contactless Payments

Raja Hasnain Anwar, Gerard DeCunha, and Muhammad Taqi Raza, University of Massachusetts Amherst

Available Media

Contactless payment cards are widely assumed to stop working past their printed expiration dates, and many stakeholders rely on this assumption for authorization and access control. This paper shows that banks enforce the expiration as a transaction policy check, rather than an intrinsic property of the card, which allows an expired card to still initiate contactless payments. We demonstrate a practical "Zombie Card" attack that makes an expired card appear unexpired, allowing successful transactions despite the card being past its printed date. We evaluate the attack across real-world transaction configurations spanning multiple EMV kernels (Visa, Mastercard, and Discover), POS terminals, merchants, and five (5) major US banks. Our results show that Visa contactless transactions are susceptible to man-in-the-middle tampering due to a lack of effective integrity protection. We further find that banks often rely on the POS terminal's decisions and skip critical security checks during transaction authorization for faster payments. Across our trials, the attack remains operational under typical in-store conditions using commodity NFC transceivers and does not require specialized hardware. Together, these findings indicate that the outcome of a "zombie card" transaction is determined by how security responsibility is divided between terminals, card manufacturers, and issuers, and by how consistently issuers enforce card lifecycle state. Based on these findings, we propose countermeasures that span kernels, issuers, and payments to ensure end-to-end transaction integrity and security.

Credible Threat Detection? Measuring Contribution Dynamics and Quality Control in a Crowdsourced Threat Detection Ecosystem

Apurva Virkud, Gang Wang, and Adam Bates, University of Illinois Urbana-Champaign

Available Media

Abstract Community exchange serves as a critical component of modern day security operations. Commercial security vendors often draw from crowdsourced intelligence and rules to power their proprietary systems. Additionally, Managed Security Service Providers (MSSPs) and Managed Detection and Response (MDR) solutions leverage both crowdsourced and commercial feeds to provide security services to organizations without their own security infrastructure. However, there is no systematic understanding of how these crowdsourced detection rules are developed or vetted. In this work, we explore the open-source and crowdsourced Sigma repository as a window into the development of threat detection rules. We find that the Sigma repository continuously iterates upon its rules, especially with tuning of false positives based on feedback from an active downstream userbase. Through an investigation of the quality control process, we observe evidence of informal rule evaluation, but also of inexperienced contributions from the broader community. Finally, we perform a comparative analysis of rules that have migrated between Sigma and other commercial Security Information and Event Management (SIEM) vendors (and vice versa) and observe divergences across platforms. Although crowdsourcing allows the community to iterate on commercial rules, we anecdotally observe that such contributions can introduce increased risk of rule evasion from the adversary. While Sigma is an active resource for vendors and the broader security community, it can benefit from more systematic and automated quality control procedures. As a first step, we introduce a tool to identify incoming duplicate rules to aid in the review process.

Facade: High-Precision Insider Threat Detection Using Deep Contextual Anomaly Detection

Alex Kantchelian, Google; Casper Neo, Palace Cybersecurity; Ryan Stevens, Hyungwon Kim, Zhaohao Fu, Sadegh Momeni, Birkett Huber, Cem Topcuoglu, Senaka Buthpitiya, Elie Bursztein, Yanis Pavlidis, Martin Cochran, and Massimiliano Poletto, Google

Available Media

Insiders with privileged access have the power to cause great harm to their organization. Even a single insider threat incident can be catastrophic, resulting in both financial losses and reputation damage. These threats are some of the most difficult to detect, as attack activity is interspersed in large volumes of legitimate activity. Although it is a serious threat, the literature is sparse aside from a few studies with various limitations, preventing their practical deployment in large scale organizations.

We present Facade (Fast and Accurate Contextual Anomaly DEtection), a high-precision, deep-learning system that has served as the last line of defense against insider threats at Google since 2018. Facade is an innovative self-supervised system that detects suspicious events by considering the context surrounding each event, including relevant facts about the user and resource involved. It is built around a new multi-modal model that is trained on corporate document access, SQL table access, and HTTP/RPC request logs. To overcome the scarcity of labeled incident data, Facade employs a novel contrastive learning strategy trained exclusively on benign activity.

Facade detects insider attackers with an extremely low false positive rate, lower than 0.01%. For single rogue events, such as the illegitimate access to a sensitive document, the false positive rate is as low as 0.0003%. To the best of our knowledge, Facade is the only proposed insider risk anomaly detection system with a false positive rate low enough for use in large corporate environments.

Concurrency Fuzzing of the Linux Kernel with eBPF

Jiacheng Xu, Zhejiang University; Dylan Wolff, Xing Yi Han, Jialin Li, and Abhik Roychoudhury, National University of Singapore

Available Media

Concurrency is indispensable for modern software systems to meet performance and scalability demands, yet concurrency bugs remain notoriously difficult to detect and reproduce. Controlled Concurrency Testing (CCT) mitigates this challenge by systematically exploring thread interleavings through scheduling control. However, existing CCT approaches for OS kernels largely rely on external enforcement mechanisms, such as custom hypervisors or invasive kernel patches, resulting in substantial overhead and limited maintainability and extensibility. In this work, we present SECT, the first kernel-native concurrency fuzzing framework that rethinks scheduling as a first-class exploration mechanism. SECT introduces a novel CCT scheduler with temporal isolation scheduling and embeds programmable scheduling policies directly into the kernel dispatch path via eBPF, enabling fine-grained control over thread interleavings without customized hypervisors or extensive kernel core modification. In addition, SECT provides a preemption-safe instrumentation mechanism for injecting scheduling points at critical kernel events and incorporates a two-phase fuzzing workflow to jointly explore both sequential and concurrent behaviors. Our evaluation demonstrates that SECT achieves 38% more branches, 57% overhead reduction and 11.4× speed-up in bug exposure compared to the state-of-the-art kernel concurrency fuzzers. Moreover, SECT discovers eight previously unknown concurrency-related bugs in the Linux kernel, five of which have already been confirmed and fixed by developers.

Bond: Constraint-Directed Fuzzing for Automated Validation of Taint Analysis Results in Linux-based IoT Firmware

Jiaqian Peng, Institute of Information Engineering, Chinese Academy of Sciences; School of Cyber Security, University of Chinese Academy of Sciences; Puzhuo Liu, Ant Group; Tsinghua University; Kai Cheng, Institute of Information Engineering, Chinese Academy of Sciences; Zhaoteng Yan, School of Cyber Security, University of Chinese Academy of Sciences; Jie Liu, Institute of Information Engineering, Chinese Academy of Sciences; Chengnian Sun, University of Waterloo; Hongsong Zhu, Institute of Information Engineering, Chinese Academy of Sciences; School of Cyber Security, University of Chinese Academy of Sciences

Available Media

Firmware vulnerabilities in IoT devices pose serious security threats, yet state-of-the-art taint analysis tools often generate large numbers of reports with limited validation. We present Bond, a directed fuzzing framework that bridges static taint analysis and dynamic vulnerability validation. Bond introduces constraint-guided input mutation by integrating three categories of constraints with six semantic types, enabling efficient exploration of paths associated with taint reports. We evaluate Bond on 19 IoT devices from 8 vendors, covering 2,776 taint reports produced by four state-of-the-art taint analyzers. Bond successfully validated 1,349 reports as real vulnerabilities, including 155 previously unknown vulnerabilities, of which 108 have been assigned CVE/PSV identifiers. On 60 known vulnerabilities, Bond achieved a 91.67% recall rate. Compared with four leading IoT fuzzers, Bond improves vulnerability validation by up to 5.5X. Ablation studies further demonstrate the effectiveness of Bond's key components and constraint extraction. These results establish Bond as a practical and effective framework for validating firmware taint analysis results.

Ira: Efficient Transaction Replay for Distributed Systems

Adithya Bhat, Harshal Bhadreshkumar Shah, and Mohsen Minaei, Visa Research

Available Media

In primary-backup replication, the consensus latency is bounded by the time for backup nodes to replay (re-execute) transactions proposed by the primary. Our key insight is that the primary, having already executed transactions, possesses knowledge of future access patterns which is the information needed for optimal replay by the backups. In this work, we present Ira, a framework to accelerate backup replay by transmitting compact hints alongside transaction batches.

We use Ethereum for our case study and present a concrete protocol, Ira-L, within our framework to improve cache management of Ethereum block execution. The primaries implementing Ira-L provide hints that consist of the working set of keys used in an Ethereum block and one byte of metadata per key indicating the table to read from, and backups use these hints for efficient block replay.

We evaluated Ira-L against the state-of-the-art Ethereum client reth over two weeks of Ethereum blocks (100,800) blocks, 24 million transactions). Our hint generation adds 10.9% overhead to primary execution time. On the backup, our hint-driven prefetching speeds up aggregate replay by 5.2× with a single prefetch thread, and by 23.6× with 16 threads. Our hints add a median of 47 KB compressed metadata per block (∼5%) of block payload).

Track 5

Security and Privacy for Web, Mobile, and Emerging Technologies 4

Session Chair: Xiaokuan Zhang, George Mason University

Dover Room

Beyond the Stars: Multimodal Detection of Scams on GitHub

Tillson Galloway, Kevin Valakuzhy, Manos Antonakakis, and Fabian Monrose, Georgia Institute of Technology

Available Media

The open-source software development ecosystem has become a prime target for phishing and supply-chain attacks, precisely because its openness encourages trust. Defenders fighting these attacks operate in an environment where adversaries generate malicious artifacts at industrial scale, evade signature-based detection through trivial mutations, and exploit delays inherent in community-driven labeling. This combination produces sparse, biased ground truth and extreme class imbalance—conditions under which conventional evaluation pipelines break down. Unsurprisingly, many detection systems that excel offline degrade sharply in real-world deployments, limiting their impact against sustained campaigns.

We present OctoWatch, a multimodal machine learning system for expanding detection of large-scale malicious source code repositories under real-world operational constraints while enabling rapid analyst pivoting through learned representations. The system fuses natural-language representations of repository metadata and file contents with structural embeddings of user–repository interactions and temporal activity statistics. Evaluated on real-world datasets—including large-scale secret-exposure attacks, prior threat-intelligence reporting, and a previously unreported phishing campaign—the approach consistently outperforms prior baselines in both precision and Matthews Correlation Coefficient (MCC). To assess operational viability, we conduct an online deployment that processes over ten million newly active repositories and identifies 14,667 total malicious repositories absent from major OSINT feeds. The system further enables discovery of coordinated campaigns, including one activity cluster associated with $5.9 million USD in cryptocurrency transactions on the Ethereum blockchain. Responsible disclosure of these findings led to the takedown of thousands of malicious repositories, demonstrating that multimodal fusion enables scalable detection of software-ecosystem abuse that systematically evades existing community-driven defenses.

SoK: Another Arms Race -- 20 Years of (Anti-)Cheating in Video Games

Philip Klostermeyer, Jan-Ulrich Holtgrave, Jacques Suray, Anne Vonderheide, Nicolas Lodzig, Niklas Busch, Alexander Krause, Jan H. Klemmer, Marten Oltrogge, and Sascha Fahl, CISPA Helmholtz Center for Information Security

Available Media

Video gaming is one of the largest sectors of the software industry. Game security is a paramount yet uniquely challenging concern. Beyond defending against conventional security threats, video games must deal with a highly motivated adversary from their own ranks: the cheater.

Cheaters seek to gain unfair advantages, undermining both the economic viability and the competitive fairness of video games. By the virtue of being an ever existing plague, the industry and related work indicate having hardships with addressing cheating. In particular, the community lacks a cohesive, systematic understanding of cheating in video games. Current literature is fragmented, focusing on isolated exploits rather than a unified architectural view of the threat landscape.

Our work addresses this critical gap by providing a comprehensive systematization of video game cheating. We systematize 187 scientific publications spanning the last two decades. Based on established knowledge systematization methodologies, we synthesize our findings into a detailed, holistic attack tree, spanning six major cheating areas with 192 nodes. We map the heterogeneous cheating strategies to the MITRE ATT&CK framework, providing a structured taxonomy of game-specific attack vectors. We enrich our work by presenting 54 safeguards for the attacks.

Motion in the Clear: Reconstructing VR User Behavior from Network Traffic

JiHo Lee, Virginia Tech; JinYi Yoon, Virginia Tech and Inha University; Taejoong Chung, Brendan David-John, and Bo Ji, Virginia Tech

Available Media

Virtual Reality (VR) applications stream high-rate head and hand motion to support real-time multi-user interaction. This motion telemetry is privacy sensitive, enabling behavioral inference such as user identification and virtual typing inference. We show that VR motion can be reconstructed from passively observed network traffic alone, even when the adversary has no application access, device compromise, or message-format knowledge. From an ecosystem measurement of 75 popular VR applications, we find that multiplayer uplink traffic is dominated by sustained, high-frequency UDP and that payload encryption is uncommon. Despite application-specific serialization and unknown encodings, motion updates retain a smooth temporal structure that remains visible in plaintext UDP payload bytes.

Leveraging this property, we present PACMO, a passive, encoding-agnostic attack that uses controlled input patterns to automatically localize motion-related packet fields and reconstruct head and hand trajectories in a black-box manner. The reconstructed motion enables high-impact downstream attacks, achieving up to 96.2% user identification accuracy and 69.4% virtual typing inference accuracy. Our results expose a systemic network-layer privacy risk in immersive systems and motivate treating motion synchronization traffic as sensitive by default.

MEPS: Privacy-preserving Edge-cloud Video Foundation Model Inference with Privacy Protectability

Siping Shi, The University of Hong Kong; Rui Lu, The Hong Kong Polytechnic University; Dan Wang, Hong Kong University of Science and Technology; Bihai Zhang, The Hong Kong Polytechnic University

Available Media

Edge-cloud video understanding systems, which leverage video foundation models on the cloud for inference to serve downstream tasks, present significant privacy risks due to the exposure of video data containing sensitive information (e.g., human faces) from edge devices to the cloud. To mitigate privacy leakage, existing methods often involve perturbing sensitive information, which can unfortunately degrade the accuracy of video understanding tasks, especially when that sensitive information is relevant to the tasks. Alternatively, encryption-based private inference offers strong privacy guarantees without accuracy loss but imposes a substantial computational burden on resource-constrained edge devices. This paper addresses these privacy challenges by observing that while many video frames contain sensitive information, only a few contain sensitive information that is truly essential for the video understanding tasks. Thus, it is unnecessary and inefficient to apply universal, computation-intensive encryption for all video frames on the edge.

We propose MEPS, a novel privacy-preserving video understanding system designed to intelligently manage the trade-off between privacy, accuracy, and efficiency. MEPS selectively encrypts video frames only when the sensitive information they contain is crucial for the downstream tasks, thus ensuring robust privacy for essential data. For the majority of frames where sensitive information is non-essential to the task, MEPS applies lightweight perturbation schemes. This approach effectively protects privacy by obscuring sensitive details without negatively impacting task accuracy or overburdening edge devices. A core component of MEPS is its ability to discern whether sensitive information within a frame is essential for video understanding. To facilitate this, we propose privacy protectability, a novel metric rooted in information theory that evaluates the degree of overlap between sensitive information and the information required for video understanding tasks. We implement a prototype of MEPS and evaluate its performance with three public real-world video trace datasets. Theoretical analysis proves MEPS has rigorous privacy guarantee. Experimental results demonstrate that MEPS can maintain the satisfied privacy protection and task accuracy, while significantly improve system efficiency, achieving a 48.8× speedup.

XGuardian: Towards Generalized, Explainable and More Effective Server-side Anti-cheat in First-Person Shooter Games

Jiayi Zhang, Chenxin Sun, and Chenxiong Qian, The University of Hong Kong

Available Media

Aim-assist cheats are the most prevalent and infamous form of cheating in First-Person Shooter (FPS) games, which help cheaters illegally reveal the opponent's location and auto-aim and shoot, and thereby pose significant threats to the game industry. Although a considerable research effort has been made to automatically detect aim-assist cheats, existing works suffer from unreliable frameworks, limited generalizability, high overhead, low detection performance, and a lack of explainability of detection results. In this paper, we propose XGuardian, a server-side generalized and explainable system for detecting aim-assist cheats to overcome these limitations. It requires only two raw data inputs, pitch and yaw, which are all FPS games' must-haves, to construct novel temporal features and describe aim trajectories, which are essential for distinguishing cheaters and normal players. XGuardian is evaluated with the latest mainstream FPS game CS2, and validates its generalizability with two different games. It achieves high detection performance and low overhead compared to prior works across different games with real-world and large-scale datasets, demonstrating wide generalizability and high effectiveness. It is able to justify its predictions and thereby shorten the manual review latency. We make XGuardian and our datasets publicly available.

VIPER Strike: Defeating Visual Reasoning CAPTCHAs via Structured Vision–Language Inference

Minfeng Qi and Dongyang He, City University of Macau; Qin Wang, CSIRO Data61; Lefeng Zhang, City University of Macau

Available Media

Visual Reasoning CAPTCHAs (VRCs) combine visual scenes with natural-language queries that demand compositional inference over objects, attributes, and spatial relations. They are increasingly deployed as a primary defense against automated bots. Existing solvers fall into two paradigms: vision-centric, which rely on template-specific detectors but fail on novel layouts, and reasoning-centric, which leverage LLMs but struggle with fine-grained visual perception. Both lack the generality needed to handle heterogeneous VRC deployments.

We present VIPER, a unified attack framework that integrates structured multi-object visual perception with adaptive LLM-based reasoning. VIPER parses visual layouts, grounds attributes to question semantics, and infers target coordinates within a modular pipeline. Evaluated on six major VRC providers (VTT, Geetest, NetEase, Dingxiang, Shumei, Xiaodun), VIPER achieves up to 93.2% success, surpassing human accuracy on most benchmarks. Compared to prior solvers, GraphNet (83.2%), Oedipus (65.8%), and the Holistic approach (89.5%), VIPER consistently outperforms all baselines. The framework further maintains robustness across alternative LLM backbones (GPT, Grok, DeepSeek, Kimi), sustaining accuracy above 90%.

To anticipate defense, we further introduce Template-Space Randomization (TSR), a lightweight strategy that perturbs linguistic templates without altering task semantics. TSR measurably reduces solver (i.e., attacker) performance. Our proposed design suggests directions for human-solvable but machine-resistant CAPTCHAs.

Send and Pretend: Exploiting Transcript Consistency Issues in End-to-End Encrypted Group Chats

Gabriel K. Gegenhuber, Interdisciplinary Transformation University (IT:U); Moritz Grefner, University of Vienna; Maximilian Günther, SBA Research; Matthäus Wininger, University of Vienna; David Schmidt, University of Vienna, UniVie Doctoral School Computer Science, and CDL AsTra; Aljosha Judmayer, University of Vienna

Available Media

End-to-end encrypted (E2EE) messaging apps are widely praised for their security and thus also used for sensitive coordination in group chats (e.g., by political decision makers). After Threema and WhatsApp, also Signal and iMessage have recently introduced polls to aid agreement processes in groups. This implicitly sets the expectation that all participants see the same outcome and thus have the same view of the conversation. This property is commonly referred to as transcript consistency (TC). In this work, we demonstrate that today's major E2EE messengers do not guarantee any form of TC for group chats, allowing a malicious group member to selectively omit, reorder, or present altered content to different recipients without triggering warnings in their user interface. We systematically investigate the extent of the problem under a malicious-participant threat model that targets the integrity of the shared transcript, or inconsistent delivery across a user's linked devices. We identify multiple equivocation vectors that range from protocol fallback paths to deliberate use of pairwise delivery channels within groups. We demonstrate concrete exploitation scenarios such as social engineering, evading moderation, and, in particular, rigging polls. Beyond these cross-service design issues, we also uncover implementation-specific behaviors with privacy implications (e.g., device OS fingerprinting). Finally, we contextualize our findings within prior transcript-consistency research and outline practical low-overhead mitigations and UI signaling strategies that can be integrated into state-of-the-art E2EE group protocols.

Relay and Betray: Exploiting Client-Side Authority in Multi-User Mixed Reality

Mutahar Ali and Habiba Farrukh, University of California, Irvine

Available Media

Multi-user mixed reality (MR) apps create shared immersive environments where users interact through embodied avatars and virtual objects in real time. These apps are increasingly used in education, workforce training, enterprise collaboration, and social experiences. To maintain immersion, multi-user MR apps minimize latency by making clients compute and broadcast state updates, such as object tracking and interaction updates. Many apps also let users upload user-generated content, such as custom avatars and 3D assets, which is distributed to other participants. This design implicitly assumes client-side trust, where each client is expected to behave honestly: report correct state updates and upload well-formed content. However, a malicious user can exploit this trust to access information that should not be perceptually or logically available to them, inject false updates, bypass moderation and safety features, or upload malicious user-generated content. While prior work has demonstrated side-channels, perception manipulation, and privacy attacks on MR systems, the security implications of client-side trust in shared MR environments remain largely unexplored.

In this paper, we present the first systematic analysis of security risks introduced by client-side trust in multi-user MR apps. We model how users and virtual objects are represented and synchronized across clients, derive security invariants that capture the conditions for safe and correct interaction, and develop a methodology to test for violations via runtime instrumentation and malicious user-generated avatars. Applying this methodology to 20 popular multi-user MR apps across different use cases and platforms, we identify five attack categories: (1) video and audio surveillance, (2) tampering of virtual objects, (3) circumvention of safety features, (4) disruption and denial-of-service, and (5) impersonation of other users. Our findings show that architectural reliance on client-side trust in multi-user MR allows a single malicious user to compromise the privacy, security, and safety of other users.

VSG-Safe: Spotting NSFW Video through Cross-Frame Evidence

Yuyang Zhang, Xudong Jiang, Yuxuan Song, and Yuxiang Sun, Wuhan University; Yihao Huang, National University of Singapore; Run Wang, Shundi Xiao, and Lina Wang, Wuhan University

Available Media

Recent advances in text-to-video (T2V) models enable high-fidelity videos that closely follow textual prompts. However, this expands practical applications while amplifying serious security and societal concerns from the automated synthesis of visual content that may be inappropriate in certain usage contexts, such as public or workplace settings, including sexual or violent content (e.g., the Grok can generate sexual videos in the "Spicy" mode). We observe that such visual content is often distributed across frames, embedded in visual entities, their attributes, and inter-entity relations. In contrast, existing moderation pipelines primarily treat video content as either individual frames or raw frame sequences, overlooking the fact that critical semantics can manifest through the combination of specific frames. This gap prevents them from reasoning across frames, confining detection to low-level visual cues, such as gore or explicit conflict, and causing frequent failures when cross-frame inference is required, including illegal activities or threats. To address these limitations, we propose leveraging scene graphs as the core intermediate semantic representation. Scene graphs naturally encode entities, their attributes, and inter-entity relationships, while also supporting reasoning over cross-frame content. Grounded on this insight, we further propose VSG-Safe, a novel scene-graph-driven framework for T2V content moderation. Concretely, our approach first extracts cross-frame content from videos to build scene graphs. With these graphs, we leverage a graph-oriented model to jointly capture entities, attributes, and inter-entity relations, enabling effective detection. To evaluate its effectiveness, we conduct extensive experiments on both SOTA benchmarks and our self-constructed video datasets. VSGSafe attains an average F1-score of 97.62%, outperforming seven baselines by 42.32% on average.