Tanya Verma, Tinfoil
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.
