American fuzzy lop – a security-oriented fuzzer

Re:American fuzzy lop – a security-oriented fuzzer by XMPPwocky on Tuesday January 28, 2020 @12:57PM (#22172037)

AFL is neat because it's conceptually simple, but practically revolutionized fuzzing, especially for non-specialists. The basic algorithm is

  def run_test_case(input):
    # return a set() of instructions that the target executes when ran on `input`
    # or throw a CrashException if the target crashes
  
  def mutate(input):
    # mess with the input- flip some bits, delete chunks, set things to 0xffffffff... randomly
    # return the mutated input

  def fuzz(initial_test_cases):
    test_cases = initial_test_cases
    coverage_seen = set()
    # collect coverage from the initial inputs
    for case in test_cases:
      coverage_seen += run_test_case(case)

    while True:
      fuzzed = mutate(random.choice(test_cases))
      try:
        new_coverage = run_test_case(fuzzed) - coverage_seen
        if new_coverage:
          # ooh, this input did something we've never seen before!
          # save it, so it can be used as a starting point
          # for even more mutation
          test_cases.add(fuzzed)
          coverage_seen += new_coverage
      except CrashException:
          # we successfully crashed the target!
          # save fuzzed off to disk or something and log a happy message

There's more to it in practice, of course- for example, run_test_case doesn't return a set of instructions, it's a bitmap / psuedo-Bloom filter of basic blocks hit. (A basic block, to a first approximation, is "a sequence of instructions that doesn't have any unusual control flow" - so if you run the first instruction in a basic block, you'll run all the rest.) And there's a fair bit of complexity involved for performance reasons.

But the core algorithm is simple enough that you can actually implement it in pure Python, for Python programs, in < 100LOC - and that implementation will find real bugs in a lot of Python libraries.

[ Parent | on HN ]

Re:American fuzzy lop – a security-oriented fuzzer by toasted_flakes on Tuesday January 28, 2020 @01:35PM (#22172557)

Nitpick, but what made AFL much more successful than other fuzzers is that it's not collecting coverage of basic blocks, but a (quantized and bloom-ish filtered) count of basic block transitions from block X to block Y.

[ Parent | on HN ]

Re:American fuzzy lop – a security-oriented fuzzer by XMPPwocky on Tuesday January 28, 2020 @01:05PM (#22172146)

And speaking of its simplicity - the original code to instrument programs involves, uh... doing this on assembly files. https://github.com/google/AFL/blob/master/afl-as.c#L284

Makes a pretty good case for "worse is better", if you like.

[ Parent | on HN ]

Re:American fuzzy lop – a security-oriented fuzzer by f- on Tuesday January 28, 2020 @02:22PM (#22173085)

(Author here)

The funniest part is that this ugly hack kept working across platforms for many years; whereas when somebody else implemented a "proper" integration with the clang / llvm API, their solution proved to be extremely fragile. The API wasn't stable between compiler versions, and because it wasn't really used much, it had all kinds of bugs, including being outright unusable at times.

Also, most distros packaged clang in a way that made it impossible to compile the plugin, because of missing or mismatched headers, missing companions tools, etc. So you had to download and rebuild the whole compiler, which took hours (and that's if you didn't get stuck in a dependency hell).

So yeah, this was very much a lesson in "worse is better".

[ Parent | on HN ]

Re:American fuzzy lop – a security-oriented fuzzer by touisteur on Tuesday January 28, 2020 @05:24PM (#22174766)

+ it allows working with GCC-based compilers (GNAT for example) and mixed C/Ada code bases.

Although the instrumentation via asm patching works well in most cases, it can break down in strange ways. See (shameless plug) : https://blog.adacore.com/running-american-fuzzy-lop-on-your-...

Amazing tool, quite extensible (adapted it to work only in memory+tmpfs without touching disk - specific corner-case... Very easily) and readable. It's also funny to scale to multiple cores and multiple machines.

[ Parent | on HN ]

Re:American fuzzy lop – a security-oriented fuzzer by jcims on Tuesday January 28, 2020 @03:33PM (#22173747)

One other benefit for me personally is that finding this little chunk of code really demystified a lot of what was happening. It’s been a few years but I still recall the aha tingles from the whole thing.

I suck at assembly but this proved to be a nice spot to hack away at it (with copious googling) to even further improve my understanding of not only the instrumentation but the effects that mutations had on program execution.

So for me worse was definitely better.

[ Parent | on HN ]

Re:American fuzzy lop – a security-oriented fuzzer by cjbprime on Tuesday January 28, 2020 @01:45PM (#22172672)

Neat! I hear it's becoming more common to use non-random mutations. Are people using something other than AFL for those, or is there a way to bolt that on to AFL?

[ Parent | on HN ]

Re:American fuzzy lop – a security-oriented fuzzer by kodablah on Tuesday January 28, 2020 @02:11PM (#22172973)

Last I reviewed, the mutations only fell through to "random havoc" after other stages were tried. So for each mutation, there are deterministic bit flips, number adds, "interesting" numbers set, dictionary uses, and only _then_ does it start the random havoc stage on the mutation. So it's not all random and you can start it with whatever value you want via dictionary or other.

(I kinda made a Java version https://github.com/cretz/javan-warty-pig that attempts to mimic this logic)

[ Parent | on HN ]