Check out the new USENIX Web site.

Home About USENIX Events Membership Publications Students
Security '05 Paper    [Security '05 Technical Program]

Finding Security Vulnerabilities in Java Applications
with Static Analysis

V. Benjamin Livshits and Monica S. Lam
Computer Science Department
Stanford University

{livshits,lam}@cs.stanford.edu

Abstract

This paper proposes a static analysis technique for detecting many recently discovered application vulnerabilities such as SQL injections, cross-site scripting, and HTTP splitting attacks. These vulnerabilities stem from unchecked input, which is widely recognized as the most common source of security vulnerabilities in Web applications. We propose a static analysis approach based on a scalable and precise points-to analysis. In our system, user-provided specifications of vulnerabilities are automatically translated into static analyzers. Our approach finds all vulnerabilities matching a specification in the statically analyzed code. Results of our static analysis are presented to the user for assessment in an auditing interface integrated within Eclipse, a popular Java development environment.
Our static analysis found 29 security vulnerabilities in nine large, popular open-source applications, with two of the vulnerabilities residing in widely-used Java libraries. In fact, all but one application in our benchmark suite had at least one vulnerability.Context sensitivity, combined with improved object naming, proved instrumental in keeping the number of false positives low. Our approach yielded very few false positives in our experiments: in fact, only one of our benchmarks suffered from false alarms.

1  Introduction

The security of Web applications has become increasingly important in the last decade. More and more Web-based enterprise applications deal with sensitive financial and medical data, which, if compromised, can cause significant downtime and millions of dollars in damages. It is crucial to protect these applications from hacker attacks.
However, the current state of application security leaves much to be desired. The 2002 Computer Crime and Security Survey conducted by the Computer Security Institute and the FBI revealed that, on a yearly basis, over half of all databases experience at least one security breach and an average episode results in close to $4 million in losses [10]. A recent penetration testing study performed by the Imperva Application Defense Center included more than 250 Web applications from e-commerce, online banking, enterprise collaboration, and supply chain management sites [54]. Their vulnerability assessment concluded that at least 92% of Web applications are vulnerable to some form of hacker attacks. Security compliance of application vendors is especially important in light of recent U.S. industry regulations such as the Sarbanes-Oxley act pertaining to information security  [4,19].
A great deal of attention has been given to network-level attacks such as port scanning, even though, about 75% of all attacks against Web servers target Web-based applications, according to a recent survey [24]. Traditional defense strategies such as firewalls do not protect against Web application attacks, as these attacks rely solely on HTTP traffic, which is usually allowed to pass through firewalls unhindered. Thus, attackers typically have a direct line to Web applications.
Many projects in the past focused on guarding against problems caused by the unsafe nature of C, such as buffer overruns and format string vulnerabilities [12,45,51]. However, in recent years, Java has emerged as the language of choice for building large complex Web-based systems, in part because of language safety features that disallow direct memory access and eliminate problems such as buffer overruns. Platforms such as J2EE (Java 2 Enterprise Edition) also promoted the adoption of Java as a language for implementing e-commerce applications such as Web stores, banking sites, etc.
A typical Web application accepts input from the user browser and interacts with a back-end database to serve user requests; J2EE libraries make these common tasks easy to code. However, despite Java language's safety, it is possible to make logical programming errors that lead to vulnerabilities such as SQL injections [1,2,14] and cross-site scripting attacks [7,22,46]. A simple programming mistake can leave a Web application vulnerable to unauthorized data access, unauthorized updates or deletion of data, and application crashes leading to denial-of-service attacks.

1.1  Causes of Vulnerabilities

Of all vulnerabilities identified in Web applications, problems caused by unchecked input are recognized as being the most common [41]. To exploit unchecked input, an attacker needs to achieve two goals:

Inject malicious data into Web applications. Common methods used include:
  • Parameter tampering: pass specially crafted malicious values in fields of HTML forms.
  • URL manipulation: use specially crafted parameters to be submitted to the Web application as part of the URL.
  • Hidden field manipulation: set hidden fields of HTML forms in Web pages to malicious values.
  • HTTP header tampering: manipulate parts of HTTP requests sent to the application.
  • Cookie poisoning: place malicious data in cookies, small files sent to Web-based applications.
Manipulate applications using malicious data. Common methods used include:
  • SQL injection: pass input containing SQL commands to a database server for execution.
  • Cross-site scripting: exploit applications that output unchecked input verbatim to trick the user into executing malicious scripts.
  • HTTP response splitting: exploit applications that output input verbatim to perform Web page defacements or Web cache poisoning attacks.
  • Path traversal: exploit unchecked user input to control which files are accessed on the server.
  • Command injection: exploit user input to execute shell commands.
These kinds of vulnerabilities are widespread in today's Web applications. A recent empirical study of vulnerabilities found that parameter tampering, SQL injection, and cross-site scripting attacks account for more than a third of all reported Web application vulnerabilities [49]. While different on the surface, all types of attacks listed above are made possible by user input that has not been (properly) validated. This set of problems is similar to those handled dynamically by the taint mode in Perl [52], even though our approach is considerably more extensible. We refer to this class of vulnerabilities as the tainted object propagation problem.

1.2  Code Auditing for Security

Many attacks described in the previous section can be detected with code auditing. Code reviews pinpoint potential vulnerabilities before an application is run. In fact, most Web application development methodologies recommend a security assessment or review step as a separate development phase after testing and before application deployment [40,41].
Code reviews, while recognized as one of the most effective defense strategies [21], are time-consuming, costly, and are therefore performed infrequently. Security auditing requires security expertise that most developers do not possess, so security reviews are often carried out by external security consultants, thus adding to the cost. In addition to this, because new security errors are often introduced as old ones are corrected, double-audits (auditing the code twice) is highly recommended. The current situation calls for better tools that help developers avoid introducing vulnerabilities during the development cycle.

1.3  Static Analysis

This paper proposes a tool based on a static analysis for finding vulnerabilities caused by unchecked input. Users of the tool can describe vulnerability patterns of interest succinctly in PQL [35], which is an easy-to-use program query language with a Java-like syntax. Our tool, as shown in Figure 1, applies user-specified queries to Java bytecode and finds all potential matches statically. The results of the analysis are integrated into Eclipse, a popular open-source Java development environment [13], making the potential vulnerabilities easy to examine and fix as part of the development process.
data/flowdiagram.png
Figure 1: Architecture of our static analysis framework.
The advantage of static analysis is that it can find all potential security violations without executing the application. The use of bytecode-level analysis obviates the need for the source code to be accessible. This is especially important since libraries whose source is unavailable are used extensively in Java applications. Our approach can be applied to other forms of bytecode such as MSIL, thereby enabling the analysis of C# code [37].
Our tool is distinctive in that it is based on a precise context-sensitive pointer analysis that has been shown to scale to large applications [55]. This combination of scalability and precision enables our analysis to find all vulnerabilities matching a specification within the portion of the code that is analyzed statically. In contrast, previous practical tools are typically unsound [6,20]. Without a precise analysis, these tools would find too many potential errors, so they only report a subset of errors that are likely to be real problems. As a result, they can miss important vulnerabilities in programs.

1.4  Contributions

A unified analysis framework. We unify multiple, seemingly diverse, recently discovered categories of security vulnerabilities in Web applications and propose an extensible tool for detecting these vulnerabilities using a sound yet practical static analysis for Java.
A powerful static analysis. Our tool is the first practical static security analysis that utilizes fully context-sensitive pointer analysis results. We improve the state of the art in pointer analysis by improving the object-naming scheme. The precision of the analysis is effective in reducing the number of false positives issued by our tool.
A simple user interface. Users of our tool can find a variety of vulnerabilities involving tainted objects by specifying them using PQL [35]. Our system provides a GUI auditing interface implemented on top of Eclipse, thus allowing users to perform security audits quickly during program development.
Experimental validation. We present a detailed experimental evaluation of our system and the static analysis approach on a set of large, widely-used open-source Java applications. We found a total of 29 security errors, including two important vulnerabilities in widely-used libraries. Eight out of nine of our benchmark applications had at least one vulnerability, and our analysis produced only 12 false positives.

1.5  Paper Organization

The rest of the paper is organized as follows. Section 2 presents a detailed overview of application-level security vulnerabilities we address. Section 3 describes our static analysis approach. Section 4 describes improvements that increase analysis precision and coverage. Section 5 describes the auditing environment our system provides. Section 6 summarizes our experimental findings. Section 7 describes related work, and Section 8 concludes.

2  Overview of Vulnerabilities

In this section we focus on a variety of security vulnerabilities in Web applications that are caused by unchecked input. According to an influential survey performed by the Open Web Application Security Project [41], unvalidated input is the number one security problem in Web applications. Many such security vulnerabilities have recently been appearing on specialized vulnerability tracking sites such as SecurityFocus and were widely publicized in the technical press [39,41]. Recent reports include SQL injections in Oracle products [31] and cross-site scripting vulnerabilities in Mozilla Firefox [30].

2.1  SQL Injection Example

Let us start with a discussion of SQL injections, one of the most well-known kinds of security vulnerabilities found in Web applications. SQL injections are caused by unchecked user input being passed to a back-end database for execution [1,2,14,29,32,47]. The hacker may embed SQL commands into the data he sends to the application, leading to unintended actions performed on the back-end database. When exploited, a SQL injection may cause unauthorized access to sensitive data, updates or deletions from the database, and even shell command execution.
A simple example of a SQL injection is shown below:
Example 1:
    HttpServletRequest request = ...;

    String userName = request.getParameter("name");

    Connection con  = ...

    String query    = "SELECT * FROM Users " +

                  " WHERE name = '" + userName + "'";

    con.execute(query);



This code snippet obtains a user name (userName) by invoking request.getParameter("name") and uses it to construct a query to be passed to a database for execution (con.execute(query)). This seemingly innocent piece of code may allow an attacker to gain access to unauthorized information: if an attacker has full control of string userName obtained from an HTTP request, he can for example set it to ' OR  1=1; -. Two dashes are used to indicate comments in the Oracle dialect of SQL, so the WHERE clause of the query effectively becomes the tautology name ="   OR  1=1. This allows the attacker to circumvent the name check and get access to all user records in the database.  []
SQL injection is but one of the vulnerabilities that can be formulated as tainted object propagation problems. In this case, the input variable userName is considered tainted. If a tainted object (the source or any other object derived from it) is passed as a parameter to con.execute (the sink), then there is a vulnerability. As discussed above, such an attack typically consists of two parts: (1) injecting malicious data into the application and (2) using the data to manipulating the application. The former corresponds to the sources of a tainted object propagation problem and the latter to the sinks. The rest of this section presents attack techniques and examples of how exploits may be created in practice.

2.2  Injecting Malicious Data

Protecting Web applications against unchecked input vulnerabilities is difficult because applications can obtain information from the user in a variety of different ways. One must check all sources of user-controlled data such as form parameters, HTTP headers, and cookie values systematically. While commonly used, client-side filtering of malicious values is not an effective defense strategy. For example, a banking application may present the user with a form containing a choice of only two account numbers; however, this restriction can be easily circumvented by saving the HTML page, editing the values in the list, and resubmitting the form. Therefore, inputs must be filtered by the Web application on the server. Note that many attacks are relatively easy to mount: an attacker needs little more than a standard Web browser to attack Web applications in most cases.

2.2.1  Parameter Tampering

The most common way for a Web application to accept parameters is through HTML forms. When a form is submitted, parameters are sent as part of an HTTP request. An attacker can easily tamper with parameters passed to a Web application by entering maliciously crafted values into text fields of HTML forms.

2.2.2  URL Tampering

For HTML forms that are submitted using the HTTP GET method, form parameters as well as their values appear as part of the URL that is accessed after the form is submitted. An attacker may directly edit the URL string, embed malicious data in it, and then access this new URL to submit malicious data to the application.
Example 2: Consider a Web page at a bank site that allows an authenticated user to select one of her accounts from a list and debit $100 from the account. When the submit button is pressed in the Web browser, the following URL is requested:
  http://www.mybank.com/myaccount?

              accountnumber=341948&debit_amount=100





However, if no additional precautions are taken by the Web application receiving this request, accessing
  http://www.mybank.com/myaccount?

              accountnumber=341948&debit_amount=-5000



may in fact increase the account balance.  []

2.2.3  Hidden Field Manipulation

Because HTTP is stateless, many Web applications use hidden fields to emulate persistence. Hidden fields are just form fields made invisible to the end-user. For example, consider an order form that includes a hidden field to store the price of items in the shopping cart:
    <input type="hidden" name="total_price"

                                    value="25.00">



A typical Web site using multiple forms, such as an online store will likely rely on hidden fields to transfer state information between pages. Unlike regular fields, hidden fields cannot be modified directly by typing values into an HTML form. However, since the hidden field is part of the page source, saving the HTML page, editing the hidden field value, and reloading the page will cause the Web application to receive the newly updated value of the hidden field.

2.2.4  HTTP Header Manipulation

HTTP headers typically remain invisible to the user and are used only by the browser and the Web server. However, some Web applications do process these headers, and attackers can inject malicious data into applications through them. While a normal Web browser will not allow forging the outgoing headers, multiple freely available tools allow a hacker to craft an HTTP request leading to an exploit [9]. Consider, for example, the Referer field, which contains the URL indicating where the request comes from. This field is commonly trusted by the Web application, but can be easily forged by an attacker. It is possible to manipulate the Referer field's value used in an error page or for redirection to mount cross-site scripting or HTTP response splitting attacks.

2.2.5  Cookie Poisoning

Cookie poisoning attacks consist of modifying a cookie, which is a small file accessible to Web applications stored on the user's computer [27]. Many Web applications use cookies to store information such as user login/password pairs and user identifiers. This information is often created and stored on the user's computer after the initial interaction with the Web application, such as visiting the application login page. Cookie poisoning is a variation of header manipulation: malicious input can be passed into applications through values stored within cookies. Because cookies are supposedly invisible to the user, cookie poisoning is often more dangerous in practice than other forms of parameter or header manipulation attacks.

2.2.6  Non-Web Input Sources

Malicious data can also be passed in as command-line parameters. This problem is not as important because typically only administrators are allowed to execute components of Web-based applications directly from the command line. However, by examining our benchmarks, we discovered that command-line utilities are often used to perform critical tasks such as initializing, cleaning, or validating a back-end database or migrating the data. Therefore, attacks against these important utilities can still be dangerous.

2.3  Exploiting Unchecked Input

Once malicious data is injected into an application, an attacker may use one of many techniques to take advantage of this data, as described below.

2.3.1  SQL Injections

SQL injections first described in Section 2.1 are caused by unchecked user input being passed to a back-end database for execution. When exploited, a SQL injection may cause a variety of consequences from leaking the structure of the back-end database to adding new users, mailing passwords to the hacker, or even executing arbitrary shell commands.
Many SQL injections can be avoided relatively easily with the use of better APIs. J2EE provides the PreparedStatement class, that allows specifying a SQL statement template with ?'s indicating statement parameters. Prepared SQL statements are precompiled, and expanded parameters never become part of executable SQL. However, not using or improperly using prepared statements still leaves plenty of room for errors.

2.3.2  Cross-site Scripting Vulnerabilities

Cross-site scripting occurs when dynamically generated Web pages display input that has not been properly validated [7,11,22,46]. An attacker may embed malicious JavaScript code into dynamically generated pages of trusted sites. When executed on the machine of a user who views the page, these scripts may hijack the user account credentials, change user settings, steal cookies, or insert unwanted content (such as ads) into the page. At the application level, echoing the application input back to the browser verbatim enables cross-site scripting.

2.3.3  HTTP Response Splitting

HTTP response splitting is a general technique that enables various new attacks including Web cache poisoning, cross-user defacement, sensitive page hijacking, as well as cross-site scripting [28]. By supplying unexpected line break CR and LF characters, an attacker can cause two HTTP responses to be generated for one maliciously constructed HTTP request. The second HTTP response may be erroneously matched with the next HTTP request. By controlling the second response, an attacker can generate a variety of issues, such as forging or poisoning Web pages on a caching proxy server. Because the proxy cache is typically shared by many users, this makes the effects of defacing a page or constructing a spoofed page to collect user data even more devastating. For HTTP splitting to be possible, the application must include unchecked input as part of the response headers sent back to the client. For example, applications that embed unchecked data in HTTP Location headers returned back to users are often vulnerable.

2.3.4  Path Traversal

Path-traversal vulnerabilities allow a hacker to access or control files outside of the intended file access path. Path-traversal attacks are normally carried out via unchecked URL input parameters, cookies, and HTTP request headers. Many Java Web applications use files to maintain an ad-hoc database and store application resources such as visual themes, images, and so on.
If an attacker has control over the specification of these file locations, then he may be able to read or remove files with sensitive data or mount a denial-of-service attack by trying to write to read-only files. Using Java security policies allows the developer to restrict access to the file system (similar to using chroot jail in Unix). However, missing or incorrect policy configuration still leaves room for errors. When used carelessly, IO operations in Java may lead to path-traversal attacks.

2.3.5  Command Injection

Command injection involves passing shell commands into the application for execution. This attack technique enables a hacker to attack the server using access rights of the application. While relatively uncommon in Web applications, especially those written in Java, this attack technique is still possible when applications carelessly use functions that execute shell commands or load dynamic libraries.

3  Static Analysis

In this section we present a static analysis that addresses the tainted object propagation problem described in Section 2.

3.1  Tainted Object Propagation

We start by defining the terminology that was informally introduced in Example 1. We define an access path as a sequence of field accesses, array index operations, or method calls separated by dots. For instance, the result of applying access path f.g to variable v is v.f.g. We denote the empty access path by ε; array indexing operations are indicated by [].
A tainted object propagation problem consists of a set of source descriptors, sink descriptors, and derivation descriptors:
  • Source descriptors of the form <m, n, p> specify ways in which user-provided data can enter the program. They consist of a source method m, parameter number n and an access path p to be applied to argument n to obtain the user-provided input. We use argument number -1 to denote the return result of a method call.
  • Sink descriptors of the form <m, n, p> specify unsafe ways in which data may be used in the program. They consist of a sink method m, argument number n, and an access path p applied to that argument.
  • Derivation descriptors of the form <m, ns, ps, nd, pd> specify how data propagates between objects in the program. They consist of a derivation method m, a source object given by argument number ns and access path ps, and a destination object given by argument number nd and access path pd. This derivation descriptor specifies that at a call to method m, the object obtained by applying pd to argument nd is derived from the object obtained by applying ps to argument ns.
In the absence of derived objects, to detect potential vulnerabilities we only need to know if a source object is used at a sink. Derivation descriptors are introduced to handle the semantics of strings in Java. Because Strings are immutable Java objects, string manipulation routines such as concatenation create brand new String objects, whose contents are based on the original String objects. Derivation descriptors are used to specify the behavior of string manipulation routines, so that taint can be explicitly passed among the String objects.
Most Java programs use built-in String libraries and can share the same set of derivation descriptors as a result. However, some Web applications use multiple String encodings such as Unicode, UTF-8, and URL encoding. If encoding and decoding routines propagate taint and are implemented using native method calls or character-level string manipulation, they also need to be specified as derivation descriptors. Sanitization routines that validate input are often implemented using character-level string manipulation. Since taint does not propagate through such routines, they should not be included in the list of derivation descriptors.
It is possible to obviate the need for manual specification with a static analysis that determines the relationship between strings passed into and returned by low-level string manipulation routines. However, such an analysis must be performed not just on the Java bytecode but on all the relevant native methods as well.
Example 3: We can formulate the problem of detecting parameter tampering attacks that result in a SQL injection as follows: the source descriptor for obtaining parameters from an HTTP request is:
<HttpServletRequest.getParameter(String),-1 ,ε>
The sink descriptor for SQL query execution is:
<Connection.executeQuery(String), 1,ε>.
To allow the use of string concatenation in the construction of query strings, we use derivation descriptors:
<StringBuffer.append(String),
1, ε, -1, ε>, and
<StringBuffer.toString(),
0, ε, -1, ε>
Due to space limitations, we show only a few descriptors here; more information about the descriptors in our experiments is available in our technical report [34].  []

Below we formally define a security violation: A source object for a source descriptor <m, n, p> is an object obtained by applying access path p to argument n of a call to m. A sink object for a sink descriptor <m, n, p> is an object obtained by applying access path p to argument n of a call to method m. Object o2 is derived from object o1, written derived(o1,o2), based on a derivation descriptor <m, ns, ps, nd, pd>, if o1 is obtained by applying ps to argument ns and o2 is obtained by applying pd to argument nd at a call to method m. An object is tainted if it is obtained by applying relation derived to a source object zero or more times. A security violation occurs if a sink object is tainted. A security violation consists of a sequence of objects o1... ok such that o1 is a source object and ok is a sink object and each object is derived from the previous one:


0 ≤ i < k : derived(oi, oi+1).
We refer to object pair <o1,ok> as a source-sink pair.

3.2  Specifications Completeness

The problem of obtaining a complete specification for a tainted object propagation problem is an important one. If a specification is incomplete, important errors will be missed even if we use a sound analysis that finds all vulnerabilities matching a specification. To come up with a list of source and sink descriptors for vulnerabilities in our experiments, we used the documentation of the relevant J2EE APIs.
Since it is relatively easy to miss relevant descriptors in the specification, we used several techniques to make our problem specification more complete. For example, to find some of the missing source methods, we instrumented the applications to find places where application code is called by the application server.
We also used a static analysis to identify tainted objects that have no other objects derived from them, and examined methods into which these objects are passed. In our experience, some of these methods turned out to be obscure derivation and sink methods missing from our initial specification, which we subsequently added.

3.3  Static Analysis

Our approach is to use a sound static analysis to find all potential violations matching a vulnerability specification given by its source, sink, and derivation descriptors. To find security violations statically, it is necessary to know what objects these descriptors may refer to, a general problem known as pointer or points-to analysis.

3.3.1  Role of Pointer Information

To illustrate the need for points-to information, we consider the task of auditing a piece of Java code for SQL injections caused by parameter tampering, as described in Example 1.
Example 4: In the code below, string param is tainted because it is returned from a source method getParameter. So is buf1, because it is derived from param in the call to append on line 6. Finally, string query is passed to sink method executeQuery.
  1   String param = req.getParameter("user");

  2

  3   StringBuffer buf1;

  4   StringBuffer buf2;

  5   ...

  6   buf1.append(param);

  7   String query = buf2.toString();

  8   con.executeQuery(query);



Unless we know that variables buf1 and buf2 may never refer to the same object, we would have to conservatively assume that they may. Since buf1 is tainted, variable query may also refer to a tainted object. Thus a conservative tool that lacks additional information about pointers will flag the call to executeQuery on line 8 as potentially unsafe.  []
An unbounded number of objects may be allocated by the program at run time, so, to compute a finite answer, the pointer analysis statically approximates dynamic program objects with a finite set of static object "names". A common approximation approach is to name an object by its allocation site, which is the line of code that allocates the object.

3.3.2  Finding Violations Statically

Points-to information enables us to find security violations statically. Points-to analysis results are represented as the relation pointsto(v, h), where v is a program variable and h is an allocation site in the program.
A static security violation is a sequence of heap allocation sites h1... hk such that
  1. There exists a variable v1 such that pointsto(v1, h1), where v1 corresponds to access path p applied to argument n of a call to method m for a source descriptor <m, n, p>.
  2. There exists a variable vk such that pointsto(vk, hk), where vk corresponds to applying access path p to argument n in a call to method m for a sink descriptor <m, n, p>.
  3. There exist variables v1,...,vk such that
    1 ≤ i < k : pointsto(vi, hi) and pointsto(vi+1,hi+1),
    where variable vi corresponds to applying ps to argument ns and vi+1 corresponds applying pd to argument nd in a call to method m for a derivation descriptor <m, ns, ps, nd,pd>.
Our static analysis is based on a context-sensitive Java points-to analysis developed by Whaley and Lam [55]. Their algorithm uses binary decision diagrams (BDDs) to efficiently represent and manipulate points-to results for exponentially many contexts in a program. They have developed a tool called bddbddb (BDD-Based Deductive DataBase) that automatically translates program analyses expressed in terms of Datalog [50] (a language used in deductive databases) into highly efficient BDD-based implementations. The results of their points-to analysis can also be accessed easily using Datalog queries. Notice that in the absence of derived objects, finding security violations can be easily done with pointer analysis alone, because pointer analysis tracks objects as they are passed into or returned from methods.
However, it is relatively easy to implement the tainted object propagation analysis using bddbddb . Constraints of a specification as given by Definition 3.6 can be translated into Datalog queries straightforwardly. Facts such as "variable v is parameter n of a call to method m" map directly into Datalog relations representing the internal representation of the Java program. The points-to results used by the constraints are also readily available as Datalog relations after pointer analysis has been run.
Because Java supports dynamic loading and classes can be dynamically generated on the fly and called reflectively, we can find vulnerabilities only in the code available to the static analysis. For reflective calls, we use a simple analysis that handles common uses of reflection to increase the size of the analyzed call graph [34].

3.3.3  Role of Pointer Analysis Precision

Pointer analysis has been the subject of much compiler research over the last two decades. Because determining what heap objects a given program variable may point to during program execution is undecidable, sound analyses compute conservative approximations of the solution. Previous points-to approaches typically trade scalability for precision, ranging from highly scalable but imprecise techniques [48] to precise approaches that have not been shown to scale [43].
In the absence of precise information about pointers, a sound tool would conclude that many objects are tainted and hence report many false positives. Therefore, many practical tools use an unsound approach to pointers, assuming that pointers are unaliased unless proven otherwise [6,20]. Such an approach, however, may miss important vulnerabilities.
Having precise points-to information can significantly reduce the number of false positives. Context sensitivity refers to the ability of an analysis to keep information from different invocation contexts of a method separate and is known to be an important feature contributing to precision. The effect of context sensitivity on analysis precision is illustrated by the example below.
       1   class DataSource {

       2       String url;

       3       DataSource(String url) {

       4           this.url = url;

       5       }

       6       String getUrl(){

       7           return this.url;

       8       }

       9       ...

      10   }

      11   String passedUrl = request.getParameter("...");

      12   DataSource ds1   = new DataSource(passedUrl);

      13   String localUrl  = "http://localhost/";

      14   DataSource ds2   = new DataSource(localUrl);

      15

      16   String s1        = ds1.getUrl();

      17   String s2        = ds2.getUrl();

Figure 2: Example showing the importance of context sensitivity.
Example 5: Consider the code snippet in Figure 2. The class DataSource acts as a wrapper for a URL string. The code creates two DataSource objects and calls getUrl on both objects. A context-insensitive analysis would merge information for calls of getUrl on lines 16 and 17. The reference this, which is considered to be argument 0 of the call, points to the object on line 12 and 14, so this.url points to either the object returned on line 11 or "http://localhost/" on line 13. As a result, both s1 and s2 will be considered tainted if we rely on context-insensitive points-to results. With a context-sensitive analysis, however, only s2 will be considered tainted.  []
While many points-to analysis approaches exist, until recently, we did not have a scalable analysis that gives a conservative yet precise answer. The context-sensitive, inclusion-based points-to analysis by Whaley and Lam is both precise and scalable [55]. It achieves scalability by using BDDs to exploit the similarities across the exponentially many calling contexts.
A call graph is a static approximation of what methods may be invoked at all method calls in the program. While there are exponentially many acyclic call paths through the call graph of a program, the compression achieved by BDDs makes it possible to efficiently represent as many as 1014 contexts. The framework we propose in this paper is the first practical static analysis tool for security to leverage the BDD-based approach. The use of BDDs has allowed us to scale our framework to programs consisting of almost 1,000 classes.

3.4  Specifying Taint Problems in PQL

While a useful formalism, source, sink, and derivation descriptors as defined in Section 3.1 are not a user-friendly way to describe security vulnerabilities. Datalog queries, while giving the user complete control, expose too much of the program's internal representation to be practical. Instead, we use PQL, a program query language. PQL serves as syntactic sugar for Datalog queries, allowing users to express vulnerability patterns in a familiar Java-like syntax; translation of tainted object propagation queries from PQL into Datalog is straightforward. PQL is a general query language capable of expressing a variety of questions about program execution. However, we only use a limited form of PQL queries to formulate tainted object propagation problems.
Due to space limitations, we summarize only the most important features of PQL here; interested readers are referred to [35] for a detailed description. A PQL query is a pattern describing a sequence of dynamic events that involves variables referring to dynamic object instances. The uses clause declares all object variables the query refers to. The matches clause specifies the sequence of events on object variables that must occur for a match. Finally, the return clause specifies the objects returned by the query whenever a set of object instances participating in the events in the matches clause is found.
Source-sink object pairs corresponding to static security violations for a given tainted object propagation problem are computed by query main in Figure 3. This query uses auxiliary queries source and sink used to define source and sink objects as well as query derived* shown in Figure 4 that captures a transitive derivation relation. Object sourceObj in main is returned by sub-query source. Object sinkObj is the result of sub-query derived^ with sourceObj used as a sub-query parameter and is also the result of sub-query sink. Therefore, sinkObj returned by query main matches all tainted objects that are also sink objects.
    

    query main()

    returns

      object Object sourceObj, sinkObj;

    matches {

        sourceObj := source();

        sinkObj   := derived*(sourceObj);

        sinkObj   := sink();

    }



Figure 3: Main query for finding source-sink pairs.
    query derived*(object Object x)

    returns

      object Object y;

    uses

      object Object temp;

    matches{

      y    := x  | 

      temp := derived(x); y := derived*(temp);

    }

Figure 4: Transitive derived relation .
Semicolons are used in PQL to indicate a sequence of events that must occur in order. Sub-query derived* defines a transitive derived relation: object y is transitively derived from object x by applying sub-query derived zero or more times. This query takes advantage of PQL's sub-query mechanism to define a transitive closure recursively. Sub-queries source, sink, and derived are specific to a particular tainted object propagation problem, as shown in the example below.
Example 6: This example describes sub-queries source, sink, and derived shown in Figure 5 that can be used to match SQL injections, such as the one described in Example 1. Usually these sub-queries are structured as a series of alternatives separated by |. The wildcard character _  is used instead of a variable name if the identity of the object to be matched is irrelevant.
    query source()

    returns

      object Object             sourceObj;

    uses

      object String[]           sourceArray;

      object HttpServletRequest req;

    matches {

          sourceObj       = req.getParameter(_) 

        | sourceObj       = req.getHeader(_)

        | sourceArray     = req.getParameterValues(_);

          sourceObj       = sourceArray[]

        |  ...

    }

    query sink()

    returns

      object Object                      sinkObj;

    uses

      object java.sql.Statement          stmt;

      objectjava.sql.Connection          con;

    matches {

          stmt.executeQuery(sinkObj)

        |  stmt.execute(sinkObj)

        |  con.prepareStatement(sinkObj)

        |  ...

    }

    query derived(objectObject x)

    returns

      object Object y;

    matches {

          y.append(x)

        |  y = _.append(x)

        |  y = new String(x)

        |  y = new StringBuffer(x)

        |  y = x.toString()

        |  y = x.substring(_,_)

        |  y = x.toString(_)

        |  ...

    }

Figure 5: PQL sub-queries for finding SQL injections.
Query source is structured as an alternation: sourceObj can be returned from a call to req.getParameter or req.getHeader for an object req of type HttpServletRequest; sourceObj may also be obtained by indexing into an array returned by a call to req.getParameterValues, etc. Query sink defines sink objects used as parameters of sink methods such as java.sql.Connection.executeQuery, etc. Query derived determines when data propagates from object x to object y. It consists of the ways in which Java strings can be derived from one another, including string concatenation, substring computation, etc.  []
As can be seen from this example, sub-queries source, sink, and derived map to source, sink, and derivation descriptors for the tainted object propagation problem. However, instead of descriptor notation for method parameters and return values, natural Java-like method invocation syntax is used.
       1     class Vector {

       2       Object[] table = new Object[1024];

       3

       4       void add(Object value){

       5           int i    = ...;

       6           // optional resizing ...

       7           table[i] = value;

       8       }

       9

      10       Object getFirst(){

      11           Object value = table[0];

      12           return value;

      13       }

      14     }

      15     String s1 = "...";

      16     Vector v1 =  new Vector();

      17     v1.add(s1);

      18     Vector v2 =  new Vector();

      19     String s2 = v2.getFirst();

Figure 6: Typical container definition and usage.

4  Precision Improvements

This section describes improvements we made to the object-naming scheme used in the original points-to analysis [55]. These improvements greatly increase the precision of the points-to results and reduce the number of false positives produced by our analysis.

4.1  Handling of Containers

Containers such as hash maps, vectors, lists, and others are a common source of imprecision in the original pointer analysis algorithm. The imprecision is due to the fact that objects are often stored in a data structure allocated inside the container class definition. As a result, the analysis cannot statically distinguish between objects stored in different containers.
Example 7: The abbreviated vector class in Figure 6 allocates an array called table on line 2 and vectors v1 and v2 share that array. As a result, the original analysis will conclude that the String object referred to by s2 retrieved from vector v2 may be the same as the String object s1 deposited in vector v1.  []
To alleviate this problem and improve the precision of the results, we create a new object name for the internally allocated data structure for every allocation site of the external container. This new name is associated with the allocation site of the underlying container object. As a result, the type of imprecision described above is eliminated and objects deposited in a container can only be retrieved from a container created at the same allocation site. In our implementation, we have applied this improved object naming to standard Java container classes including HashMap, HashTable, and LinkedList.

4.2  Handling of String Routines

Another set of methods that requires better object naming is Java string manipulation routines. Methods such as String.toLowerCase() allocate String objects that are subsequently returned. With the default object-naming scheme, all the allocated strings are considered tainted if such a method is ever invoked on a tainted string.
We alleviate this problem by giving unique names to results returned by string manipulation routines at different call sites. We currently apply this object naming improvement to Java standard libraries only. As explained in Section 6.4, imprecise object naming was responsible for all the 12 false positives produced by our analysis.

5  Auditing Environment

The static analysis described in the previous two sections forms the basis of our security-auditing tool for Java applications. The tool allows a user to specify security patterns to detect. User-provided specifications are expressed as PQL queries, as described in Section 3.4. These queries are automatically translated into Datalog queries, which are subsequently resolved using bddbddb .
To help the user with the task of examining violation reports, our provides an intuitive GUI interface. The interface is built on top of Eclipse, a popular open-source Java development environment. As a result, a Java programmer can assess the security of his application, often without leaving the development environment used to create the application in the first place.
A typical auditing session involves applying the analysis to the application and then exporting the results into Eclipse for review. Our Eclipse plugin allows the user to easily examine each vulnerability by navigating among the objects involved in it. Clicking on each object allows the user to navigate through the code displayed in the text editor in the top portion of the screen.
out6.png
Figure 7: Tracking a SQL injection vulnerability in the Eclipse GUI plugin. Objects involved in the vulnerability trace are shown at the bottom.
Example 8: An example of using the Eclipse GUI is shown in Figure 7. The bottom portion of the screen lists all potential security vulnerabilities reported by our analysis. One of them, a SQL injection caused by non-Web input is expanded to show all the objects involved in the vulnerability. The source object on line 76 of JDBCDatabaseExport.java is passed to derived objects using derivation methods StringBuffer.append and StringBuffer.toString until it reaches the sink object constructed and used on line 170 of the same file. Line 170, which contains a call to Connection.prepareStatement, is highlighted in the Java text editor shown on top of the screen.  []

6  Experimental Results

In this section we summarize the experiments we performed and described the security violations we found. We start out by describing our benchmark applications and experimental setup, describe some representative vulnerabilities found by our analysis, and analyze the impact of analysis features on precision.

6.1  Benchmark Applications

While there is a fair number of commercial and open-source tools available for testing Web application security, there are no established benchmarks for comparing tools' effectiveness. The task of finding suitable benchmarks for our experiments was especially complicated by the fact that most Web-based applications are proprietary software, whose vendors are understandably reluctant to reveal their code, not to mention the vulnerabilities found. At the same time, we did not want to focus on artificial micro-benchmarks or student projects that lack the complexities inherent in real applications. We focused on a set of large, representative open-source Web-based J2EE applications, most of which are available on SourceForge.
The benchmark applications are briefly described below. jboard, blueblog, blojsom, personalblog, snipsnap, pebble, and roller are Web-based bulletin board and blogging applications. webgoat is a J2EE application designed by the Open Web Application Security Project [40,41] as a test case and a teaching tool for Web application security. Finally, road2hibernate is a test program developed for hibernate, a popular object persistence library.
Applications were selected from among J2EE-based open-source projects on SourceForge solely on the basis of their size and popularity. Other than webgoat, which we knew had intentional security flaws, we had no prior knowledge as to whether the applications had security vulnerabilities. Most of our benchmark applications are used widely: roller is used on dozens of sites including prominent ones such as blogs.sun.com. snipsnap has more than 50,000 downloads according to its authors. road2hibernate is a wrapper around hibernate, a highly popular object persistence library that is used by multiple large projects, including a news aggregator and a portal. personalblog has more than 3,000 downloads according to SourceForge statistics. Finally, blojsom was adopted as a blogging solution for the Apple Tiger Weblog Server.
Figure 8 summarizes information about our benchmark applications. Notice that the traditional lines-of-code metric is somewhat misleading in the case of applications that use large libraries. Many of these benchmarks depend on massive libraries, so, while the application code may be small, the full size of the application executed at runtime is quite large. An extreme case is road2hibernate, which is a small 140-line stub program designed to exercise the hibernate object persistence library; however, the total number of analyzed classes for road2hibernate exceeded 800. A better measure is given in the last column of Figure 8, which shows the total number of classes in each application's call graph.

Version File Line Analyzed
Benchmark number count count classes
jboard 0.30 90 17542 264
blueblog 1.0 32 4191 306
webgoat 0.9 77 19440 349
blojsom 1.9.6 61 14448 428
personalblog 1.2.6 39 5591 611
snipsnap 1.0-BETA-1 445 36745 653
road2hibernate 2.1.4 2 140 867
pebble 1.6-beta1 333 36544 889
roller 0.9.9 276 52089 989
Total 1355 186730 5356
Figure 8: Summary of information about the benchmarks. Applications are sorted by the total number of analyzed classes.

6.2  Experimental Setup

The implementation of our system is based on the joeq Java compiler and analysis framework. In our system we use a translator from PQL to Datalog [35] and the bddbddb program analysis tool [55] to find security violations. We applied static analysis to look for all tainted object propagation problems described in this paper, and we used a total of 28 source,  18 sink, and 29 derivation descriptors in our experiments. The derivation descriptors correspond to methods in classes such as String, StringBuffer, StringTokenizer, etc. Source and sink descriptors correspond to methods declared in 19 different J2EE classes, as is further described in [34].

Pre- Points-to analysis Taint analysis
Context sensitivity proces- xxxx
Improved naming sing xx xx
jboard 37 871210 14121614
blueblog 39 1381510 17142116
webgoat 57 453011890 6966106101
blojsom 60 18132516 24213027
personalblog 173 1072830332 62501959
snipsnap 193 583314247 194154160105
road2hibernate 247 1864026843 734416158
pebble 177 583511749 150140136100
roller 362 22655733103 19683338129

Figure 9: Summary of times, in seconds, it takes to perform preprocessing, points-to, and taint analysis for each analysis variation. Analysis variations have either context sensitivity or improved object naming enabled, as indicated by x signs in the header row.
We used four different variations of our static analysis, obtained by either enabling or disabling context sensitivity and improved object naming. Analysis times for the variations are listed in Figure 9. Running times shown in the table are obtained on an Opteron 150 machine with 4 GB of memory running Linux. The first section of the table shows the times to pre-process the application to create relations accepted by the pointer analysis; the second shows points-to analysis times; the last presents times for the tainted object propagation analysis.
It should be noted that the taint analysis times often decrease as the analysis precision increases. Contrary to intuition, we actually pay less for a more precise analysis. Imprecise answers are big and therefore take a long time to compute and represent. In fact, the context-insensitive analysis with default object naming runs significantly slower on the largest benchmarks than the most precise analysis. The most precise analysis version takes a total of less than 10 minutes on the largest application; we believe that this is acceptable given the quality of the results the analysis produces.

6.3  Vulnerabilities Discovered

The static analysis described in this paper reports a total of 41 potential security violations in our nine benchmarks, out of which 29 turn out to be security errors, while 12 are false positives. All but one of the benchmarks had at least one security vulnerability. Moreover, except for errors in webgoat and one HTTP splitting vulnerability in snipsnap [16], none of these security errors had been reported before.

6.3.1  Validating the Errors We Found

Not all security errors found by static analysis or code reviews are necessarily exploitable in practice. The error may not correspond to a path that can be taken dynamically, or it may not be possible to construct meaningful malicious input. Exploits may also be ruled out because of the particular configuration of the application, but configurations may change over time, potentially making exploits possible. For example, a SQL injection that may not work on one database may become exploitable when the application is deployed with a database system that does not perform sufficient input checking. Furthermore, virtually all static errors we found can be fixed easily by modifying several lines of Java source code, so there is generally no reason not to fix them in practice.
After we ran our analysis, we manually examined all the errors reported to make sure they represent security errors. Since our knowledge of the applications was not sufficient to ascertain that the errors we found were exploitable, to gain additional assurance, we reported the errors to program maintainers. We only reported to application maintainers only those errors found in the application code rather than general libraries over which the maintainer had no control. Almost all errors we reported to program maintainers were confirmed, resulting in more that a dozen code fixes.

Sources Sinks Tainted objects Reported warnings False positivesErrors
Context sensitivity xxxxxx
Improved object namingxxxxxx
jboard 1 6 268 23 2 2 0 0 0 0 0 0 0 0 0
blueblog 6 12 17 17 17 17 1 1 1 1 0 0 0 0 1
webgoat 13 59 1166 201 903 157 51 7 51 6 45 1 45 0 6
blojsom 27 18 368 203 197 112 48 4 26 2 46 2 24 0 2
personalblog 25 31 2066 1023 1685 426 460 275 370 2 458 273 368 0 2
snipsnap 155 100 1168 791 897 456 732 93 513 27 717 78 498 12 15
road2hibernate 1 33 2150 843 1641 385 18 12 16 1 17 11 15 0 1
pebble 132 70 1403 621 957 255 427 211 193 1 426 210 192 0 1
roller 32 64 2367 504 1923 151 378 12 261 1 377 11 260 0 1
Total 392 393 10973 4226 8222 1961 2115 615 1431 41 2086 586 1402 12 29

data/paircomp.png
Figure 10: (a) Summary of data on the number of tainted objects, reported security violations, and false positives for each analysis version. Enabled analysis features are indicated by x signs in the header row. (b) Comparison of the number of tainted objects for each version of the analysis.
Because webgoat is an artificial application designed to contain bugs, we did not report the errors we found in it. Instead, we dynamically confirmed some of the statically detected errors by running webgoat, as well as a few other benchmarks, on a local server and creating actual exploits.
It is important to point out that our current analysis ignores control flow. Without analyzing the predicates, our analysis may not realize that a program has checked its input, so some of the reported vulnerabilities may turn out to be false positives. However, our analysis shows all the steps involved in propagating taint from a source to a sink, thus allowing the user to check if the vulnerabilities found are exploitable.
Many Web-based application perform some form of input checking. However, as in the case of the vulnerabilities we found in snipsnap, it is common that some checks are missed. It is surprising that our analysis did not generate any false warnings due to the lack of predicate analysis, even though many of the applications we analyze include checks on user input. Two security errors in blojsom flagged by our analysis deserve special mention. The user-provided input was in fact checked, but the validation checks were too lax, leaving room for exploits. Since the sanitization routine in blojsom was implemented using string operations as opposed to direct character manipulation, our analysis detected the flow of taint from the routine's input to its output. To prove the vulnerability to the application maintainer, we created an exploit that circumvented all the checks in the validation routine, thus making path-traversal vulnerabilities possible. Note that if the sanitation was properly implemented, our analysis would have generated some false positives in this case.

6.3.2  Classification of Errors

This section presents a classification of all the errors we found. A summary of our experimental results is presented in Figure 10(a). Columns 2 and 3 list the number of source and sink objects for each benchmark. It should be noted that the number of sources and sinks for all of these applications is quite large, which suggests that security auditing these applications is time-consuming, because the time a manual security code review takes is roughly proportional to the number of sources and sinks that need to be considered. The table also shows the number of vulnerability reports, the number of false positives, and the number of errors for each analysis version.
Figure 11 presents a classification of the 29 security vulnerabilities we found grouped by the type of the source in the table rows and the sink in table columns. For example, the cell in row 4, column 1 indicates that there were 2 potential SQL injection attacks caused by non-Web sources, one in snipsnap and another in road2hibernate.
Overall, parameter manipulation was the most common technique to inject malicious data (13 cases) and HTTP splitting was the most popular exploitation technique (11 cases). Many HTTP splitting vulnerabilities are due to an unsafe programming idiom where the application redirects the user's browser to a page whose URL is user-provided as the following example from snipsnap demonstrates:
    response.sendRedirect(

            request.getParameter("referer"));



Most of the vulnerabilities we discovered are in application code as opposed to libraries. While errors in application code may result from simple coding mistakes made by programmers unaware of security issues, one would expect library code to generally be better tested and more secure. Errors in libraries expose all applications using the library to attack. Despite this fact, we have managed to find two attack vectors in libraries: one in a commonly used Java library hibernate and another in the J2EE implementation. While a total of 29 security errors were found, because the same vulnerability vector in J2EE is present in four different benchmarks, they actually corresponded to 26 unique vulnerabilities.

SQL injections HTTP splitting Cross-site scripting Path traversal Total
Header manip. 0 snipsnap = 6 blueblog: 1, webgoat: 1, pebble: 1, roller: 1 = 4 0 10
Parameter manip. webgoat: 4, personalblog: 2 = 6 snipsnap = 5 0 blojsom = 2 13
Cookie poisoning webgoat = 10 0 0 1
Non-Web inputs snipsnap : 1, road2hibernate: 1 = 20 0 snipsnap = 35
Total 9 11 4 529

Figure 11: Classification of vulnerabilities we found. Each cell corresponds to a combination of a source type (in rows) and sink type (in columns).

6.3.3  SQL Injection Vector in hibernate

We start by describing a vulnerability vector found in hibernate, an open-source object-persistence library commonly used in Java applications as a lightweight back-end database. hibernate provides the functionality of saving program data structures to disk and loading them at a later time. It also allows applications to search through the data stored in a hibernate database. Three programs in our benchmark suite, personalblog, road2hibernate, and snipsnap, use hibernate to store user data.
We have discovered an attack vector in code pertaining to the search functionality in hibernate, version 2.1.4. The implementation of method Session.find retrieves objects from a hibernate database by passing its input string argument through a sequence of calls to a SQL execute statement. As a result, all invocations of Session.find with unsafe data, such as the two errors we found in personalblog, may suffer from SQL injections. A few other public methods such as iterate and delete also turn out to be attack vectors. Our findings highlight the importance of securing commonly used software components in order to protect their clients.

6.3.4  Cross-site Tracing Attacks

Analysis of webgoat and several other applications revealed a previously unknown vulnerability in core J2EE libraries, which are used by thousands of Java applications. This vulnerability pertains to the TRACE method specified in the HTTP protocol. TRACE is used to echo the contents of an HTTP request back to the client for debugging purposes. However, the contents of user-provided headers are sent back verbatim, thus enabling cross-site scripting attacks.
In fact, this variation of cross-site scripting caused by a vulnerability in HTTP protocol specification was discovered before, although the fact that it was present in J2EE was not previously announced. This type of attack has been dubbed cross-site tracing and it is responsible for CERT vulnerabilities 244729, 711843, and 728563. Because this behavior is specified by the HTTP protocol, there is no easy way to fix this problem at the source level. General recommendations for avoiding cross-site tracing include disabling TRACE functionality on the server or disabling client-side scripting [18].

6.4  Analysis Features and False Positives

The version of our analysis that employs both context sensitivity and improved object naming described in Section 4 achieves very precise results, as measured by the number of false positives. In this section we examine the contribution of each feature of our static analysis approach to the precision of our results. We also explain the causes of the remaining 12 false positives reported by the most precise analysis version. To analyze the importance of each analysis feature, we examined the number of false positives as well as the number of tainted objects reported by each variation of the analysis. Just like false positives, tainted objects provide a useful metric for analysis precision: as the analysis becomes more precise, the number of objects deemed to be tainted decreases.
Figure 10(a) summarizes the results for the four different analysis versions. The first part of the table shows the number of tainted objects reported by the analysis. The second part of the table shows the number of reported security violations. The third part of the table summarizes the number of false positives. Finally, the last column provides the number of real errors detected for each benchmark. Figure 10(b) provides a graphical representation of the number of tainted objects for different analysis variations. Below we summarize our observations.
Context sensitivity combined with improved object naming achieves a very low number of false positives. In fact, the number of false positives was 0 for all applications but snipsnap. For snipsnap, the number of false positives was reduced more than 50-fold compared to the context-insensitive analysis version with no naming improvements. Similarly, not counting the small program jboard, the most precise version on average reported 5 times fewer tainted objects than the least precise. Moreover, the number of tainted objects dropped more that 15-fold in the case of roller, our largest benchmark.
To achieve a low false-positive rate, both context sensitivity and improved object naming are necessary. The number of false positives remains high for most programs when only one of these analysis features is used. One way to interpret the importance of context sensitivity is that the right selection of object "names" in pointer analysis allows context sensitivity to produce precise results. While it is widely recognized in the compiler community that special treatment of containers is necessary for precision, improved object naming alone is not generally sufficient to completely eliminate the false positives.
All 12 of the false positives reported by the most precise version for our analysis were located in snipsnap and were caused by insufficient precision of the default allocation site-based object-naming scheme. The default naming caused an allocation site in snipsnap to be conservatively considered tainted because a tainted object could propagate to that allocation site. The allocation site in question is located within StringWriter.toString(), a JDK function similar to String.toLowerCase() that returns a tainted String only if the underlying StringWriter is constructed from a tainted string. Our analysis conservatively concluded that the return result of this method may be tainted, causing a vulnerability to be reported, where none can occur at runtime. We should mention that all the false positives in snipsnap are eliminated by creating a new object name at every call to, StringWriter.toString(), which is achieved with a one-line change to the pointer analysis specification.

7  Related Work

In this section, we first discuss penetration testing and runtime monitoring, two of the most commonly used approaches for finding vulnerabilities besides manual code reviews. We also review the relevant literature on static analysis for improving software security.

7.1  Penetration Testing

Current practical solutions for detecting Web application security problems generally fall into the realm of penetration testing [3,5,15,36,44]. Penetration testing involves attempting to exploit vulnerabilities in a Web application or crashing it by coming up with a set of appropriate malicious input values. Penetration reports usually include a list of identified vulnerabilities [25]. However, this approach is incomplete. A penetration test can usually reveal only a small sample of all possible security risks in a system without identifying the parts of the system that have not been adequately tested. Generally, there are no standards that define which tests to run and which inputs to try. In most cases this approach is not effective and considerable program knowledge is needed to find application-level security errors successfully.

7.2  Runtime Monitoring

A variety of both free and commercial runtime monitoring tools for evaluating Web application security are available. Proxies intercept HTTP and HTTPS data between the server and the client, so that data, including cookies and form fields, can be examined and modified, and resubmitted to the application [9,42]. Commercial application-level firewalls available from NetContinuum, Imperva, Watchfire, and other companies take this concept further by creating a model of valid interactions between the user and the application and warning about violations of this model. Some application-level firewalls are based on signatures that guard against known types of attacks. The white-listing approach specifies what the valid inputs are; however, maintaining the rules for white-listing is challenging. In contrast, our technique can prevent security errors before they have a chance to manifest themselves.

7.3  Static Analysis Approaches

A good overview of static analysis approaches applied to security problems is provided in  [8]. Simple lexical approaches employed by scanning tools such as ITS4 and RATS use a set of predefined patterns to identify potentially dangerous areas of a program [56]. While a significant improvement on Unix grep, these tools, however, have no knowledge of how data propagates throughout the program and cannot be used to automatically and fully solve taint-style problems.
A few projects use path-sensitive analysis to find errors in C and C++ programs [6,20,33]. While capable of addressing taint-style problems, these tools rely on an unsound approach to pointers and may therefore miss some errors. The WebSSARI project uses combined unsound static and dynamic analysis in the context of analyzing PHP programs [23]. WebSSARI has successfully been applied to find many SQL injection and cross-site scripting vulnerabilities in PHP code.
An analysis approach that uses type qualifiers has been proven successful in finding security errors in C for the problems of detecting format string violations and user/kernel bugs [26,45]. Context sensitivity significantly reduces the rate of false positives encountered with this technique; however, it is unclear how scalable the context-sensitive approach is.
Much of the work in information-flow analysis uses a type-checking approach, as exemplified by JFlow [38]. The compiler reads a program containing labeled types and, in checking the types, ensures that the program cannot contain improper information flow at runtime. The security type system in such a language enforces information-flow policies. The annotation effort, however, may be prohibitively expensive in practice. In addition to explicit information flows our approach addresses, JFlow also deals with implicit information flows.
Static analysis has been applied to analyzing SQL statements constructed in Java programs that may lead to SQL injection vulnerabilities [17,53]. That work analyzes strings that represent SQL statements to check for potential type violations and tautologies. This approach assumes that a flow graph representing how string values can propagate through the program has been constructed a priori from points-to analysis results. However, since accurate pointer information is necessary to construct an accurate flow graph, it is unclear whether this technique can achieve the scalability and precision needed to detect errors in large systems.

8  Conclusions

In this paper we showed how a general class of security errors in Java applications can be formulated as instances of the general tainted object propagation problem, which involves finding all sink objects derivable from source objects via a set of given derivation rules. We developed a precise and scalable analysis for this problem based on a precise context-sensitive pointer alias analysis and introduced extensions to the handling of strings and containers to further improve the precision. Our approach finds all vulnerabilities matching the specification within the statically analyzed code. Note, however, that errors may be missed if the user-provided specification is incomplete.
We formulated a variety of widespread vulnerabilities including SQL injections, cross-site scripting, HTTP splitting attacks, and other types of vulnerabilities as tainted object propagation problems. Our experimental results showed that our analysis is an effective practical tool for finding security vulnerabilities. We were able to find a total of 29 security errors, and all but one of our nine large real-life benchmark applications were vulnerable. Two vulnerabilities were located in commonly used libraries, thus subjecting applications using the libraries to potential vulnerabilities. Most of the security errors we reported were confirmed as exploitable vulnerabilities by their maintainers, resulting in more than a dozen code fixes. The analysis reported false positives for only one application. We determined that the false warnings reported can be eliminated with improved object naming.

9  Acknowledgements

We are grateful to Michael Martin for his help with PQL and dynamic validation of some of the vulnerabilities we found and to John Whaley for his support with the bddbddb tool and the joeq framework. We thank our paper shepherd R. Sekar, whose insightful comments helped improve this paper considerably. We thank the benchmark application maintainers for responding to our bug reports. We thank Amit Klein for providing detailed clarifications about Web application vulnerabilities and Ramesh Chandra, Chris Unkel, and Ted Kremenek and the anonymous paper reviewers for providing additional helpful comments. Finally, this material is based upon work supported by the National Science Foundation under Grant No. 0326227.


References

[1]
C. Anley. Advanced SQL injection in SQL Server applications. http://www.nextgenss.com/papers/advanced_sql_injection.pdf, 2002.
[2]
C. Anley. (more) advanced SQL injection. http://www.nextgenss.com/papers/more_advanced_sql_injection.pdf, 2002.
[3]
B. Arkin, S. Stender, and G. McGraw. Software penetration testing. IEEE Security and Privacy, 3(1):84-87, 2005.
[4]
K. Beaver. Achieving Sarbanes-Oxley compliance for Web applications through security testing. http://www.spidynamics.com/support/whitepapers/WI_SOXwhitepaper.pdf, 2003.
[5]
B. Buege, R. Layman, and A. Taylor. Hacking Exposed: J2EE and Java: Developing Secure Applications with Java Technology. McGraw-Hill/Osborne, 2002.
[6]
W. R. Bush, J. D. Pincus, and D. J. Sielaff. A static analyzer for finding dynamic programming errors. Software - Practice and Experience (SPE), 30:775-802, 2000.
[7]
CGI Security. The cross-site scripting FAQ. http://www.cgisecurity.net/articles/xss-faq.shtml.
[8]
B. Chess and G. McGraw. Static analysis for security. IEEE Security and Privacy, 2(6):76-79, 2004.
[9]
Chinotec Technologies. Paros-a tool for Web application security assessment. http://www.parosproxy.org, 2004.
[10]
Computer Security Institute. Computer crime and security survey. http://www.gocsi.com/press/20020407.jhtml?_requestid=195148, 2002.
[11]
S. Cook. A Web developer’s guide to cross-site scripting. http://www.giac.org/practical/GSEC/Steve_Cook_GSEC.pdf, 2003.
[12]
C. Cowan, C. Pu, D. Maier, J. Walpole, P. Bakke, S. Beattie, A. Grier, P. Wagle, Q. Zhang, and H. Hinton. StackGuard: Automatic adaptive detection and prevention of buffer-overflow attacks. In Proceedings of the 7th USENIX Security Conference, pages 63-78, January 1998.
[13]
J. D'Anjou, S. Fairbrother, D. Kehn, J. Kellerman, and P. McCarthy. Java Developer's Guide to Eclipse. Addison-Wesley Professional, 2004.
[14]
S. Friedl. SQL injection attacks by example. http://www.unixwiz.net/techtips/sql-injection.html, 2004.
[15]
D. Geer and J. Harthorne. Penetration testing: A duet. http://www.acsac.org/2002/papers/geer.pdf, 2002.
[16]
Gentoo Linux Security Advisory. SnipSnap: HTTP response splitting. http://www.gentoo.org/security/en/glsa/glsa-200409-23.xml, 2004.
[17]
C. Gould, Z. Su, and P. Devanbu. Static checking of dynamically generated queries in database applications. In Proceedings of the 26th International Conference on Software Engineering, pages 645-654, 2004.
[18]
J. Grossman. Cross-site tracing (XST): The new techniques and emerging threats to bypass current Web security measures using TRACE and XSS. http://www.cgisecurity.com/whitehat-mirror/WhitePaper_screen.pdf, 2003.
[19]
J. Grossman. WASC activities and U.S. Web application security trends. http://www.whitehatsec.com/presentations/WASC_WASF_1.02.pdf, 2004.
[20]
S. Hallem, B. Chelf, Y. Xie, and D. Engler. A system and language for building system-specific, static analyses. In Proceedings of the ACM SIGPLAN 2002 Conference on Programming language Design and Implementation, pages 69-82, 2002.
[21]
M. Howard and D. LeBlanc. Writing Secure Code. Microsoft Press, 2001.
[22]
D. Hu. Preventing cross-site scripting vulnerability. http://www.giac.org/practical/GSEC/Deyu_Hu_GSEC.pdf, 2004.
[23]
Y.-W. Huang, F. Yu, C. Hang, C.-H. Tsai, D.-T. Lee, and S.-Y. Kuo. Securing Web application code by static analysis and runtime protection. In Proceedings of the 13th conference on World Wide Web, pages 40-52, 2004.
[24]
G. Hulme. New software may improve application security. http://www.informationweek.com/story/IWK20010209S0003, 2001.
[25]
Imperva, Inc. SuperVeda penetration test. http://www.imperva.com/download.asp?id=3.
[26]
R. Johnson and D. Wagner. Finding user/kernel pointer bugs with type inference. In Proceedings of the 2004 Usenix Security Conference, pages 119–-134, 2004.
[27]
A. Klein. Hacking Web applications using cookie poisoning. http://www.cgisecurity.com/lib/CookiePoisoningByline.pdf, 2002.
[28]
A. Klein. Divide and conquer: HTTP response splitting, Web cache poisoning attacks, and related topics. http://www.packetstormsecurity.org/papers/general/whitepaper_httpresponse.pdf, 2004.
[29]
S. Kost. An introduction to SQL injection attacks for Oracle developers. http://www.net-security.org/dl/articles/IntegrigyIntrotoSQLInjectionAttacks.pdf, 2004.
[30]
M. Krax. Mozilla foundation security advisory 2005-38. http://www.mozilla.org/security/announce/mfsa2005-38.html, 2005.
[31]
D. Litchfield. Oracle multiple PL/SQL injection vulnerabilities. http://www.securityfocus.com/archive/1/385333/2004-12-20/2004-12-26/0, 2003.
[32]
D. Litchfield. SQL Server Security. McGraw-Hill Osborne Media, 2003.
[33]
V. B. Livshits and M. S. Lam. Tracking pointers with path and context sensitivity for bug detection in C programs. In Proceedings of the ACM SIGSOFT Symposium on the Foundations of Software Engineering, pages 317-326, Sept. 2003.
[34]
V. B. Livshits and M. S. Lam. Detecting security vulnerabilities in Java applications with static analysis. Technical report. Stanford University. http://suif.stanford.edu/~livshits/papers/tr/webappsec_tr.pdf, 2005.
[35]
M. Martin, V. B. Livshits, and M. S. Lam. Finding application errors using PQL: a program query language (to be published). In Proceedings of the ACM Conference on Object-Oriented Programming, Systems, Languages, and Applications (OOPSLA), Oct. 2005.
[36]
J. Melbourne and D. Jorm. Penetration testing for Web applications. http://www.securityfocus.com/infocus/1704, 2003.
[37]
J. S. Miller, S. Ragsdale, and J. Miller. The Common Language Infrastructure Annotated Standard. Addison-Wesley Professional, 2003.
[38]
A. C. Myers. JFlow: practical mostly-static information flow control. In Proceedings of the 26th ACM SIGPLAN-SIGACT Symposium on Principles of Programming Languages, pages 228-241, Jan. 1999.
[39]
NetContinuum, Inc. The 21 primary classes of Web application threats. https://www.netcontinuum.com/securityCentral/TopThreatTypes/index.cfm, 2004.
[40]
Open Web Application Security Project. A guide to building secure Web applications. http://voxel.dl.sourceforge.net/sourceforge/owasp/OWASPGuideV1.1.pdf, 2004.
[41]
Open Web Application Security Project. The ten most critical Web application security vulnerabilities. http://umn.dl.sourceforge.net/sourceforge/owasp/OWASPTopTen2004.pdf, 2004.
[42]
Open Web Application Security Project. WebScarab. http://www.owasp.org/software/webscarab.html, 2004.
[43]
S. Sagiv, T. Reps, and R. Wilhelm. Parametric shape analysis via 3-valued logic. In Proceedings of the 26th ACM Symposium on Principles of Programming Languages, pages 105-118, Jan. 1999.
[44]
J. Scambray and M. Shema. Web Applications (Hacking Exposed). Addison-Wesley Professional, 2002.
[45]
U. Shankar, K. Talwar, J. S. Foster, and D. Wagner. Detecting format string vulnerabilities with type qualifiers. In Proceedings of the 2001 Usenix Security Conference, pages 201-220, Aug. 2001.
[46]
K. Spett. Cross-site scripting: are your Web applications vulnerable. http://www.spidynamics.com/support/whitepapers/SPIcross-sitescripting.pdf, 2002.
[47]
K. Spett. SQL injection: Are your Web applications vulnerable? http://downloads.securityfocus.com/library/SQLInjectionWhitePaper.pdf, 2002.
[48]
B. Steensgaard. Points-to analysis in almost linear time. In Proceedings of the 23th ACM Symposium on Principles of Programming Languages, pages 32-41, Jan. 1996.
[49]
M. Surf and A. Shulman. How safe is it out there? http://www.imperva.com/download.asp?id=23, 2004.
[50]
J. D. Ullman. Principles of Database and Knowledge-Base Systems. Computer Science Press, Rockville, Md., volume II edition, 1989.
[51]
D. Wagner, J. Foster, E. Brewer, and A. Aiken. A first step towards automated detection of buffer overrun vulnerabilities. In Proceedings of Network and Distributed Systems Security Symposium, pages 3-17, Feb. 2000.
[52]
L. Wall, T. Christiansen, and R. Schwartz. Programming Perl. O'Reilly and Associates, Sebastopol, CA, 1996.
[53]
G. Wassermann and Z. Su. An analysis framework for security in Web applications. In Proceedings of the Specification and Verification of Component-Based Systems Workshop, Oct. 2004.
[54]
WebCohort, Inc. Only 10% of Web applications are secured against common hacking techniques. http://www.imperva.com/company/news/2004-feb-02.html, 2004.
[55]
J. Whaley and M. S. Lam. Cloning-based context-sensitive pointer alias analysis using binary decision diagrams. In Proceedings of the ACM SIGPLAN 2004 conference on Programming Language Design and Implementation, pages 131-144, June 2004.
[56]
J. Wilander and M. Kamkar. A comparison of publicly available tools for static intrusion prevention. In Proceedings of 7th Nordic Workshop on Secure IT Systems, Nov. 2002.

This paper was originally published in the Proceedings of the 14th USENIX Security Symposium,
July 31–August 5, 2005, Baltimore, MD

Last changed: 3 Aug. 2005 ch
Security '05 Technical Program
Security '05 Home
USENIX home