Computing Applications Practice

SAGE: Whitebox Fuzzing For Security Testing

SAGE has had a remarkable impact at Microsoft.
  1. Introduction
  2. Hunting for "Million-Dollar" Bugs
  3. Introducing Whitebox Fuzzing
  4. SAGE
  5. SAGE Architecture
  6. An Example
  7. Impact of SAGE
  8. Conclusion
  9. Acknowledgments
  10. References
  11. Authors
  12. Figures
sample jpg image of the earth from outer space
This picture is displayed on your screen after a jpg parser (typically part of your operating system) has read the image data, decoded it, created new data structures with the decoded data, and passed those to the graphics card in your computer.

back to top  

Most Communications readers might think of “program verification research” as mostly theoretical with little impact on the world at large. Think again. If you are reading these lines on a PC running some form of Windows (like over 93% of PC users—that is, more than one billion people), then you have been affected by this line of work—without knowing it, which is precisely the way we want it to be.

Every second Tuesday of every month, also known as “Patch Tuesday,” Microsoft releases a list of security bulletins and associated security patches to be deployed on hundreds of millions of machines worldwide. Each security bulletin costs Microsoft and its users millions of dollars. If a monthly security update costs you $0.001 (one tenth of one cent) in just electricity or loss of productivity, then this number multiplied by one billion people is $1 million. Of course, if malware were spreading on your machine, possibly leaking some of your private data, then that might cost you much more than $0.001. This is why we strongly encourage you to apply those pesky security updates.

Many security vulnerabilities are a result of programming errors in code for parsing files and packets that are transmitted over the Internet. For example, Microsoft Windows includes parsers for hundreds of file formats.

If you are reading this article on a computer, then the picture shown in Figure 1 is displayed on your screen after a jpg parser (typically part of your operating system) has read the image data, decoded it, created new data structures with the decoded data, and passed those to the graphics card in your computer. If the code implementing that jpg parser contains a bug such as a buffer overflow that can be triggered by a corrupted jpg image, then the execution of this jpg parser on your computer could potentially be hijacked to execute some other code, possibly malicious and hidden in the jpg data itself.

This is just one example of a possible security vulnerability and attack scenario. The security bugs discussed throughout the rest of this article are mostly buffer overflows.

Back to Top

Hunting for “Million-Dollar” Bugs

Today, hackers find security vulnerabilities in software products using two primary methods. The first is code inspection of binaries (with a good disassembler, binary code is like source code).

The second is blackbox fuzzing, a form of blackbox random testing, which randomly mutates well-formed program inputs and then tests the program with those modified inputs,3 hoping to trigger a bug such as a buffer overflow. In some cases, grammars are used to generate the well-formed inputs. This also allows encoding application-specific knowledge and test-generation heuristics.

Blackbox fuzzing is a simple yet effective technique for finding security vulnerabilities in software. Thousands of security bugs have been found this way. At Microsoft, fuzzing is mandatory for every untrusted interface of every product, as prescribed in the Security Development Lifecycle,7 which documents recommendations on how to develop secure software.

Although blackbox fuzzing can be remarkably effective, its limitations are well known. For example, the then branch of the conditional statement in


has only 1 in 232 chances of being exercised if the input variable x has a randomly chosen 32-bit value. This intuitively explains why blackbox fuzzing usually provides low code coverage and can miss security bugs.

Back to Top

Introducing Whitebox Fuzzing

A few years ago, we started developing an alternative to blackbox fuzzing, called whitebox fuzzing.5 It builds upon recent advances in systematic dynamic test generation4 and extends its scope from unit testing to whole-program security testing.

Starting with a well-formed input, whitebox fuzzing consists of symbolically executing the program under test dynamically, gathering constraints on inputs from conditional branches encountered along the execution. The collected constraints are then systematically negated and solved with a constraint solver, whose solutions are mapped to new inputs that exercise different program execution paths. This process is repeated using novel search techniques that attempt to sweep through all (in practice, many) feasible execution paths of the program while checking simultaneously many properties using a runtime checker (such as Purify, Valgrind, or AppVerifier).

For example, symbolic execution of the previous program fragment with an initial value 0 for the input variable x takes the else branch of the conditional statement and generates the path constraint x+3 ≠ 13. Once this constraint is negated and solved, it yields x = 10, providing a new input that causes the program to follow the then branch of the conditional statement. This allows us to exercise and test additional code for security bugs, even without specific knowledge of the input format. Furthermore, this approach automatically discovers and tests “corner cases” where programmers may fail to allocate memory or manipulate buffers properly, leading to security vulnerabilities.

In theory, systematic dynamic test generation can lead to full program path coverage, that is, program verification. In practice, however, the search is typically incomplete both because the number of execution paths in the program under test is huge, and because symbolic execution, constraint generation, and constraint solving can be imprecise due to complex program statements (pointer manipulations and floating-point operations, among others), calls to external operating-system and library functions, and large numbers of constraints that cannot all be solved perfectly in a reasonable amount of time. Therefore, we are forced to explore practical trade-offs.

Back to Top


Whitebox fuzzing was first implemented in the tool SAGE, short for Scalable Automated Guided Execution.5 Because SAGE targets large applications where a single execution may contain hundreds of millions of instructions, symbolic execution is its slowest component. Therefore, SAGE implements a novel directed-search algorithm—dubbed generational search—that maximizes the number of new input tests generated from each symbolic execution. Given a path constraint, all the constraints in that path are systematically negated one by one, placed in a conjunction with the prefix of the path constraint leading to it, and attempted to be solved by a constraint solver. This way, a single symbolic execution can generate thousands of new tests. (In contrast, a standard depth-first or breadth-first search would negate only the last or first constraint in each path constraint and generate at most one new test per symbolic execution.)

SAGE has had a remarkable impact at Microsoft. It combines and extends program analysis, testing, verification, model checking, and automated theorem-proving techniques that have been developed over many years.

The program shown in Figure 2 takes four bytes as input and contains an error when the value of the variable cnt is greater than or equal to four. Starting with some initial input good, SAGE runs this program while performing a symbolic execution dynamically. Since the program path taken during this first run is formed by all the else branches in the program, the path constraint for this initial run is the conjunction of constraints i0 ≠ b, i1 ≠ a, i2 ≠ d and i3 ≠ !. Each of these constraints is negated one by one, placed in a conjunction with the prefix of the path constraint leading to it, and then solved with a constraint solver. In this case, all four constraints are solvable, leading to four new test inputs. Figure 2 also shows the set of all feasible program paths for the function top. The leftmost path represents the initial run of the program and is labeled 0 for Generation 0. Four Generation 1 inputs are obtained by systematically negating and solving each constraint in the Generation 0 path constraint. By repeating this process, all paths are eventually enumerated for this example. In practice, the search is typically incomplete.

SAGE was the first tool to perform dynamic symbolic execution at the x86 binary level. It is implemented on top of the trace replay infrastructure TruScan,8 which consumes trace files generated by the iDNA framework1 and virtually re-executes the recorded runs. TruScan offers several features that substantially simplify symbolic execution, including instruction decoding, providing an interface to program symbol information, monitoring various input/output system calls, keeping track of heap and stack frame allocations, and tracking the flow of data through the program structures. Thanks to offline tracing, constraint generation in SAGE is completely deterministic because it works with an execution trace that captures the outcome of all nondeterministic events encountered during the recorded run. Working at the x86 binary level allows SAGE to be used on any program regardless of its source language or build process. It also ensures that “what you fuzz is what you ship,” as compilers can perform source-code changes that may affect security.

SAGE uses several optimizations that are crucial for dealing with huge execution traces. For example, a single symbolic execution of Excel with 45,000 input bytes executes nearly one billion x86 instructions. To scale to such execution traces, SAGE uses several techniques to improve the speed and memory usage of constraint generation: symbolic-expression caching ensures that structurally equivalent symbolic terms are mapped to the same physical object; unrelated constraint elimination reduces the size of constraint solver queries by removing the constraints that do not share symbolic variables with the negated constraint; local constraint caching skips a constraint if it has already been added to the path constraint; flip count limit establishes the maximum number of times a constraint generated from a particular program branch can be flipped; using a cheap syntactic check, constraint subsumption eliminates constraints logically implied by other constraints injected at the same program branch (mostly likely resulting from successive iterations of an input-dependent loop).

Back to Top

SAGE Architecture

The high-level architecture of SAGE is depicted in Figure 3. Given one (or more) initial input Input0, SAGE starts by running the program under test with AppVerifier to see if this initial input triggers a bug. If not, SAGE then collects the list of unique program instructions executed during this run. Next, SAGE symbolically executes the program with that input and generates a path constraint, characterizing the current program execution with a conjunction of input constraints.

Then, implementing a generational search, all the constraints in that path constraint are negated one by one, placed in a conjunction with the prefix of the path constraint leading to it, and attempted to be solved by a constraint solver (we currently use the Z3 SMT solver2). All satisfiable constraints are mapped to N new inputs, that are tested and ranked according to incremental instruction coverage. For example, if executing the program with new Input1 discovers 100 new instructions, then Input1 gets a score of 100, and so on. Next, the new input with the highest score is selected to go through the (expensive) symbolic execution task, and the cycle is repeated, possibly forever. Note that all the SAGE tasks can be executed in parallel on a multicore machine or even on a set of machines.

Building a system such as SAGE poses many other challenges: how to recover from imprecision in symbolic execution, how to check many properties together efficiently, how to leverage grammars (when available) for complex input formats, how to deal with path explosion, how to reason precisely about pointers, how to deal with floating-point instructions and input-dependent loops. Space constraints prevent us from discussing these challenges here, but the authors’ Web pages provide access to other papers addressing these issues.

Back to Top

An Example

On April 3, 2007, Microsoft released an out-of-band critical security patch (MS07-017) for code that parses ANI-format animated cursors. The vulnerability was originally reported to Microsoft in December 2006 by Alex Sotirov of Determina Security Research, then made public after exploit code appeared in the wild.9 It was only the third such out-of-band patch released by Microsoft since January 2006, indicating the seriousness of the bug. The Microsoft SDL Policy Weblog stated that extensive blackbox fuzzing of this code failed to uncover the bug and that existing static-analysis tools were not capable of finding the bug without excessive false positives.6

SAGE, in contrast, synthesized a new input file exhibiting the bug within hours of starting from a well-formed ANI file, despite having no knowledge of the ANI format. A seed file was picked arbitrarily from a library of well-formed ANI files, and SAGE was run on a small test program that called user32.dll to parse ANI files. The initial run generated a path constraint with 341 branch constraints after executing 1,279,939 total x86 instructions over 10,072 symbolic input bytes. SAGE then created a crashing ANI file after 7 hours 36 minutes and 7,706 test cases, using one core of a 2GHz AMD Opteron 270 dual-core processor running 32-bit Windows Vista with 4GB of RAM.

Back to Top

Impact of SAGE

Since 2007, SAGE has discovered many security-related bugs in many large Microsoft applications, including image processors, media players, file decoders, and document parsers. Notably, SAGE found approximately one-third of all the bugs discovered by file fuzzing during the development of Microsoft’s Windows 7. Because SAGE is typically run last, those bugs were missed by everything else, including static program analysis and blackbox fuzzing.

Finding all these bugs has saved Microsoft millions of dollars as well as saved world time and energy, by avoiding expensive security patches to more than one billion PCs. The software running on your PC has been affected by SAGE.

Since 2008, SAGE has been running 24/7 on approximately 100-plus machines/cores automatically fuzzing hundreds of applications in Microsoft security testing labs. This is more than 300 machine-years and the largest computational usage ever for any Satisfiability Modulo Theories (SMT) solver, with more than one billion constraints processed to date.

SAGE is so effective at finding bugs that, for the first time, we faced “bug triage” issues with dynamic test generation. We believe this effectiveness comes from being able to fuzz large applications (not just small units as previously done with dynamic test generation), which in turn allows us to find bugs resulting from problems across multiple components. SAGE is also easy to deploy, thanks to x86 binary analysis, and it is fully automatic. SAGE is now used daily in various groups at Microsoft.

Back to Top


SAGE has had a remarkable impact at Microsoft. It combines and extends program analysis, testing, verification, model checking, and automated theorem-proving techniques that have been developed over many years.

Which is best in practice—blackbox or whitebox fuzzing? Both offer different cost/precision trade-offs. Blackbox is simple, lightweight, easy, and fast but can yield limited code coverage. White-box is smarter but more complex.

Which approach is more effective at finding bugs? It depends. If an application has never been fuzzed, any form of fuzzing is likely to find bugs, and simple blackbox fuzzing is a good start. Once the low-hanging bugs are gone, however, fuzzing for the next bugs has to become smarter. Then it is time to use whitebox fuzzing and/or user-provided guidance, for example, using an input grammar.

The bottom line? In practice, use both. We do at Microsoft.

Back to Top


Many people across different groups at Microsoft have contributed to SAGE’s success; more than space allows here. To review this detailed list, please visit

q stamp of ACM Queue Related articles

Black Box Debugging
James A. Whittaker, Herbert H. Thompson

Security – Problem Solved?
John Viega

Open vs. Closed: Which Source is More Secure?
Richard Ford

Back to Top

Back to Top

Back to Top


F1 Figure 1. A sample jpg image.

F2 Figure 2. Example of program (left) and its search space (right) with the value of cnt at the end of each run.

F3 Figure 3. Architecture of SAGE.

Back to top

    1. Bhansali, S., Chen, W., De Jong, S., Edwards, A. and Drinic, M. Framework for instruction-level tracing and analysis of programs. In Proceedings of the 2nd International Conference on Virtual Execution Environments (2006).

    2. de Moura, L. and Bjorner, N. Z3: An efficient SMT solver. In Proceedings of Tools and Algorithms for the Construction and Analysis of Systems; Lecture Notes in Computer Science. Springer-Verlag, 2008, 337–340.

    3. Forrester, J.E. and Miller, B.P. An empirical study of the robustness of Windows NT applications using random testing. In Proceedings of the 4th Usenix Windows System Symposium (Seattle, WA, Aug. 2000).

    4. Godefroid, P., Klarlund, N. and Sen, K. DART: Directed Automated Random Testing. In Proceedings of Programming Language Design and Implementation (2005), 213–223.

    5. Godefroid, P., Levin, M. Y. and Molnar, D. Automated whitebox fuzz testing. In Proceedings of Network and Distributed Systems Security, (2008), 151–166.

    6. Howard, M. Lessons learned from the animated cursor security bug;

    7. Howard, M. and Lipner, S. The Security Development Lifecycle. Microsoft Press, 2006.

    8. Narayanasamy, S., Wang, Z., Tigani, J., Edwards, A. and Calder, B. Automatically classifying benign and harmful data races using replay analysis. In Programming Languages Design and Implementation (2007).

    9. Sotirov, A. Windows animated cursor stack overflow vulnerability, 2007;

Join the Discussion (0)

Become a Member or Sign In to Post a Comment

The Latest from CACM

Shape the Future of Computing

ACM encourages its members to take a direct hand in shaping the future of the association. There are more ways than ever to get involved.

Get Involved

Communications of the ACM (CACM) is now a fully Open Access publication.

By opening CACM to the world, we hope to increase engagement among the broader computer science community and encourage non-members to discover the rich resources ACM has to offer.

Learn More