Automated Whitebox Fuzz Testing Patrice Godefroid (Microsoft Research)‏ Michael Y. Levin (Microsoft Center for Software Excellence)‏ David Molnar (UC-Berkeley, visiting MSR) Presented by Yuyan Bao

Acknowledgement This presentation is extended and modified from The presentation by David Molnar Automated Whitebox Fuzz Testing The presentation by Corina Păsăreanu (Kestrel) Symbolic Execution for Model Checking and Testing

Agenda Fuzz Testing Symbolic Execution Dynamic Testing Generation Generational Search SAGE Architecture Conclusion Contribution Weakness Improvement Traditionally fuzz testing tools

Fuzz Testing An effective technique for finding security vulnerabilities in software Apply invalid, unexpected, or random data to the inputs of a program. If the program fails(crashing, access violation exception), the defects can be noted. Traditionally fuzz testing tools

Whitebox Fuzzing This paper present an alternative whitebox fuzz testing approach inspired by recent advances in symbolic execution and dynamic test generation Run the code with some initial well-formed input Collect constraints on input with symbolic execution Generate new constraints (by negating constraints one by one) Solve constraints with constraint solver Synthesize new inputs Just to give a brief overview, whitebox fuzzing is divided into several steps.

White Box Testing and Symbolic Execution A method for deriving test cases which satisfy a given path Performed as a simulation of the computation on the path Initial path function = Identity function, Initial path condition = true Each vertex on the path induce a symbolic composition on the path function and a logical constraint on the path condition: If an assignment was made: If a conditional decision was made: path condition path condition branch condition Output: path function and path condition for the given path Slide from “White Box Testing and Symbolic Execution” by Michael Beder White Box Testing and Symbolic Execution

Symbolic Execution Symbolic execution tree: Code: Not reachable FALSE! (PC=“path condition”) - “Simulate” the code using symbolic values instead of program numeric data Symbolic execution tree: x:X,y:Y PC: true PC: X>Y PC: X<=Y true false x:X+Y,y:Y x:X+Y,y:X x:Y,y:X PC: X>Y  Y-X>0 PC:X>Y  Y-X<=0 FALSE! Not reachable int x, y; if (x > y) { x = x + y; y = x - y; x = x - y; if (x – y > 0) assert (false); }

Dynamic Test Generation Executing the program starting with some initial inputs Performing a dynamic symbolic execution to collect constraints on inputs Use a constraint solver to infer variants of the previous inputs in order to steer the next executions

Dynamic Test Generation void top(char input[4]) { int cnt = 0; if (input[0] == ‘b’) cnt++; if (input[1] == ‘a’) cnt++; if (input[2] == ‘d’) cnt++; if (input[3] == ‘!’) cnt++; if (cnt >= 3) crash(); } input = “good” Point out this is a dynamic technique. Consider the program shown. This program takes 4 bytes as input and contains an error when the value of the variable cnt is greater than or equal to 3 at the end of the function top. We start running the function top with the initial 4-letters string “good”. Running the program with random values for the 4 input bytes is unlikely to discovery the error: a probability of about Traditional fuzz testing is difficult to generate input values that will drive program through all its possible execution paths. 9

Dynamic Test Generation void top(char input[4]) { int cnt = 0; if (input[0] == ‘b’) cnt++; if (input[1] == ‘a’) cnt++; if (input[2] == ‘d’) cnt++; if (input[3] == ‘!’) cnt++; if (cnt >= 3) crash(); } input = “good” I0 != ‘b’ I1 != ‘a’ I2 != ‘d’ I3 != ‘!’ Collect constraints from trace Create new constraints Solve new constraints  new input. 10

Depth-First Search good I0 != ‘b’ I1 != ‘a’ void top(char input[4]) { int cnt = 0; if (input[0] == ‘b’) cnt++; if (input[1] == ‘a’) cnt++; if (input[2] == ‘d’) cnt++; if (input[3] == ‘!’) cnt++; if (cnt >= 3) crash(); } I2 != ‘d’ I0 != ‘b’ I3 != ‘!’ I1 != ‘a’ I2 != ‘d’ We collect constraints from trace path that represents all the input vectors that drive the program through the path. I3 != ‘!’ good

Depth-First Search good goo! void top(char input[4]) { int cnt = 0; if (input[0] == ‘b’) cnt++; if (input[1] == ‘a’) cnt++; if (input[2] == ‘d’) cnt++; if (input[3] == ‘!’) cnt++; if (cnt >= 3) crash(); } I0 != ‘b’ I1 != ‘a’ I2 != ‘d’ To force the program through a different path constraint, one can negate the last predicate of the current path constraint. And “goo!” is a solution to this path constraint. I3 == ‘!’ good goo!

Practical limitation good godd void top(char input[4]) { int cnt = 0; if (input[0] == ‘b’) cnt++; if (input[1] == ‘a’) cnt++; if (input[2] == ‘d’) cnt++; if (input[3] == ‘!’) cnt++; if (cnt >= 3) crash(); } I0 != ‘b’ I1 != ‘a’ I2 == ‘d’ No matter Depth-First Search or Breadth-First Search, every time only one constraint is expanded. However realistic programs always have large inputs and deep paths. This paper presents a new search algorithm to address path explosion limitation I3 != ‘!’ good godd Every time only one constraint is expanded, low efficiency!

Generational search BFS with a heuristic to maximize block coverage Score returns the number of new blocks covered Slide from “Symbolic Execution” by Kevin Wallace, CSE504 2010-04-28

Key Idea: One Trace, Many Tests Generational Search Key Idea: One Trace, Many Tests Office 2007 application: Time to gather constraints: 25m30s Tainted branches/trace: ~1000 Time per branch to solve, generate new test, check for crashes: ~1s Therefore, solve+check all branches for each trace!

Key Idea: One Trace, Many Tests Generational Search Key Idea: One Trace, Many Tests bood void top(char input[4]) { int cnt = 0; if (input[0] == ‘b’) cnt++; if (input[1] == ‘a’) cnt++; if (input[2] == ‘d’) cnt++; if (input[3] == ‘!’) cnt++; if (cnt >= 3) crash(); } gaod I0 == ‘b’ godd I1 == ‘a’ I2 == ‘d’ Staring with an initial input and initial path constraint , the new search algorithm will attempt to expand all constraints, instead of just the last one with a depth-first search, or the first one with a breadth-first search. I3 == ‘!’ good goo! “Generation 1” test cases

The Generational Search Space void top(char input[4]) { int cnt = 0; if (input[0] == ‘b’) cnt++; if (input[1] == ‘a’) cnt++; if (input[2] == ‘d’) cnt++; if (input[3] == ‘!’) cnt++; if (cnt >= 3) crash(); } Consider again the program with initial input 4 letters string good. From this parent run, a generational search generates four children whose leafs are labeled with 1. Indeed those four paths each correspond to negating one constraint in the original path constraint. Each of those first generation execution paths can be expanded to generate second-generation children. By repeating this process, all feasible execution paths of the function top are eventually generate exactly once. good 1 goo! 1 godd 2 god! 1 gaod 2 gao! 2 gadd 1 bood 2 boo! 2 bodd 2 baod gad! bod! baod badd bad!

Task: SymbolicExecutor Constraints (Disolver) SAGE Architecture (Scalable Automated Guided Execution) Input0 Trace File Constraints Task: Tester Check for Crashes (AppVerifier)‏ Task: Tracer Trace Program (Nirvana)‏ Task: CoverageCollector Gather Constraints (Truscan)‏ Task: SymbolicExecutor Solve Constraints (Disolver) The generational search algorithm has been implemented in a new tool named SAGE, which stands for… SAGE can test any file-reading program running on windows by treating bytes read from files as symbolic inputs. Another key novelty of SAGE is that it performs symbolic execution of program traces at the x86 binary level Input1 Input2 … InputN

Initial Experiences with SAGE Since 1st MS internal release in April’07: dozens of new security bugs found (most missed by blackbox fuzzers, static analysis)‏ Apps: image processors, media players, file decoders,… Many bugs found rated as “security critical, severity 1, priority 1” Now used by several teams regularly as part of QA process SAGE now is used by several teams in Microsoft, many high security priority bugs found by SAGE and most were missed by blackbox fuzzing efforts.

Experiments ANI Parsing - MS07-017 RIFF...ACONLIST B...INFOINAM.... 3D Blue Alternat e v1.1..IART.... ................ 1996..anih$...$. ..rate.......... ..........seq .. ..LIST....framic on......... .. RIFF...ACONB B...INFOINAM.... 3D Blue Alternat e v1.1..IART.... ................ 1996..anih$...$. ..rate.......... ..........seq .. ..anih....framic on......... .. Only 1 in 232 chance at random! In contrast,SAGE can generate a crash starting from a well-formed specific file, having no knowledge of the file format. On the left, an ASCII rendering of init well-formed input file. On the right, the SAGE- generated crash for MS07-017. SAGE test case changes the LIST to an additional anih record. It takes more than 7 hours and near 8000 test case are generated. Initial input generate : 341 branch constraints, more than 1 million total instructions. Initial input Crashing test case Initial input : 341 branch constraints, 1,279,939 total instructions. Crash test cases: 7706 test cases, 7hrs 36 mins

Initial Experiments #Instructions and Input size largest seen so far App Tested #Tests Mean Depth Mean #Instr. Mean Size ANI 11468 178 2,066,087 5,400 Media 1 6890 73 3,409,376 65,536 Media 2 1045 1100 271,432,489 27,335 Media 3 2266 608 54,644,652 30,833 Media 4 909 883 133,685,240 22,209 Compression 1527 65 480,435 634 Office 2007 3008 6502 923,731,248 45,064 Clarify depth is in path constraint size. Clarify instructions are post-file-read. Size is variable because of different input-dependent paths. 21

Different Crashes From Different Initial Input Files 100 zero bytes Bug ID seed1 seed2 seed3 seed4 seed5 seed6 seed7 1867196225 X 2031962117 612334691 1061959981 1212954973 1011628381 842674295 1246509355 1527393075 1277839407 1951025690 Stack hash on left. Includes address of faulting instruction. All crashes are unhandled second-chance exceptions 5.9 crashes per machine hour. Does not include bogus-2,3,4,5 – didn’t find crashes Other bogus files found 0 bugs. Media 1: 60 machine-hours, 44,598 total tests, 357 crashes, 12 bugs 22

Conclusion Blackbox vs. Whitebox Fuzzing Cost/precision tradeoffs Blackbox is lightweight, easy and fast, but poor coverage Whitebox is smarter, but complex and slower Recent “semi-whitebox” approaches Less smart but more lightweight: Flayer (taint-flow analysis, may generate false alarms), Bunny-the-fuzzer (taint-flow, source-based, heuristics to fuzz based on input usage), autodafe, etc. Which is more effective at finding bugs? It depends… Many apps are so buggy, any form of fuzzing finds bugs! Once low-hanging bugs are gone, fuzzing must become smarter: use whitebox and/or user-provided guidance (grammars, etc.)‏ Bottom-line: in practice, use both!

Contribution A novel search algorithm with a coverage-maximizing heuristic It performs symbolic execution of program traces at the x86 binary level Tests larger applications than previously done in dynamic test generation

Weakness Gathering initial inputs The class of inputs characterized by each symbolic execution is determined by the dependence of the program’s control flow on its inputs Not a general solution Only for x86 windows applications Only focus on file-reading applications Nirvana simulates a given processor’s architecture

Improvement Portability Translate collected traces into an intermediate language Reproduce bugs on different platforms without re-simulating the program Better way to find initial inputs efficiently

Thank you! Questions?