CPython bytecode instrumentation and forkserver tools for fuzzing python code using AFL.
The tools in this repository enable coverage-guided fuzzing of pure python and mixed python/c code using American Fuzzy Lop (even better, AFL++).
There are three main parts to this:
- A bytecode rewriter using a technique inspired by inspired by Ned Batchelder's "wicked hack"
detailed at https://nedbatchelder.com/blog/200804/wicked_hack_python_bytecode_tracing.html.
In this case, the rewriter identifies "basic blocks" in the python bytecode and abuses the
codeobject'slnotab(line-number table) to mark each basic block as a new "line". These new "lines" are what trigger CPython's line-level trace hooks. The result of this being that we can get our trace hook executed on every new basic block. - A minimal & fast tracehook written in C, tallying visited locations to sysv shared memory.
- A basic forkserver implementation.
Preparing code for fuzzing involves a couple of steps. The first thing that should happen in
the python process is a call to install_rewriter(). It's important that this is done very
early as any modules that are imported before this will not be properly instrumented.
from cpytraceafl.rewriter import install_rewriter
install_rewriter()install_rewriter() can optionally be provided with a selector controlling which code objects
are instrumented and to what degree.
Following this, modules can be imported as normal and will be instrumented by the monkeypatched
compile functions. It's usually a good idea to initialize the test environment next,
performing as many setup procedures as possible before the input file is read. This may
include doing an initial run of the function under test to ensure any internal imports or caches
are set up. This is because we want to minimize work that has to be done post-fork - any work
done now only has to be done once,
After calling
from cpytraceafl import fuzz_from_here
fuzz_from_here()the fork() will have been made and tracing started. You now simply read your input file and
call your function under test.
Examples for fuzzing some common packages are provided in examples/.
As for hooking this script up to AFL, I tend to use the included dummy-afl-qemu-trace shim script to fool AFL's QEmu mode into communicating directly with the python process.
As of version 0.4.0, cpytraceafl can gather trace information from C extension modules that
have been compiled with AFL instrumentation (e.g. using llvm_mode). This means that it can
be used to seamlessly fuzz projects which have a mix of python and C "speedups". This is
important not only because a lot of python format-parsing packages use this approach, but
because issues revealed in native code are far more likely to have security implications.
Including instrumented native code requires a little more care when preparing a target for
fuzzing. For instance, it's important to ensure the cpytraceafl.tracehook module has been
imported and it has had its set_map_start(...) function provided with a valid memory
area before any instrumented extension modules are loaded. This is because simply loading an
instrumented native module will cause it to attempt to log its execution trace somewhere.
The example pillow_pcx_example.py demonstrates a fuzzing target taking the necessary precautions into account.
It's possible that you're only interested in tracing the native code, using cpytraceafl
just as a driver, in which case you can omit the early install_rewriter() call and all
the weirdness involved with that.
cpytraceafl-regex is a companion,
re-replacement regex implementation with added instrumentation that should aid AFL in
generating examples that pass regular expressions used in the target code, or
exercise them in interesting ways. Without this, AFL will just see regular expressions
as a black box that will act as a barrier to path exploration.
cpytraceafl has been used to find:
- Pillow: CVE-2020-10177, CVE-2020-10378, CVE-2020-10379, CVE-2020-10994, CVE-2020-11538.
- bsdiff4: CVE-2020-15904
- asyncpg: CVE-2020-17446
- clickhouse-driver: CVE-2020-26759
Well, yes and no. My experience has been that fuzzing python code is simply "a bit different" from fuzzing native code - you tend to be looking for different things. In terms of raw speed, fuzzing python is certainly not fast, but iteration rates I tend to work with aren't completely dissimilar to what I'm used to getting with AFL's Qemu mode (of course, no two fuzzing targets are really directly comparable).
Because of the memory-safe nature of pure python code, it's also more uncommon for issues uncovered through fuzzing to be security issues - logical flaws in parsing tend to lead to unexpected/unhandled exceptions. So it's still a rather useful tool in simply looking for bugs. It can be used, for example, to generate a corpus of example inputs for your test suite which exercise a large amount of the code.
However, note that while pure python code may be memory safe, as soon as you start using
the C api, Cython, or even start playing with the ctypes module, it is not.
From a rigorous academic stance, and for some uses, possibly not - you've got to keep in mind
that half the bytecode instructions could result in calls out to more arbitrary python or
(uninstrumented) native code that could have arbitrary side effects. But for our needs it works
well enough (recall that AFL coverage analysis is robust to random instrumentation
sites being omitted through AFL_INST_RATIO or AFL_INST_LIBS).
Absolutely it does. Don't use instrumented programs to debug problematic cases - use it to generate problematic inputs. Analyze them with instrumentation turned off.
Looks like you're trying to import an (instrumented) native extension module before the
cpytraceafl.tracehook module has been loaded (which is what provides that symbol).
You probably also need to provide cpytraceafl.tracehook.set_map_start(...) with a valid
writeable memory area before the import. Assuming you're not interested in the trace associated
with the import process, this can just be a dummy which you later discard. I'd recommend either
using an mmap object or sysv_ipc.SharedMemory. When fuzz_from_here() is called, this will
be replaced with right one.
It's also possible the instrumented module was built with a different AFL MAP_SIZE_POW2 from
that in cpytraceafl.MAP_SIZE_BITS.
No, you can use your normal distribution-installed python. If you're just looking at fuzzing pure python, you don't need to even think about building any binaries with funny compilers.
You may be interested in building c/c++/cython-based modules or their underlying native libraries with instrumentation if that's what you're trying to fuzz, but I suspect using a natively-instrumented cpython would be quite complicated and extremely slow.
I have tended to use tcmalloc's debugging modes with TCMALLOC_PAGE_FENCE and
TCMALLOC_PAGE_FENCE_NEVER_RECLAIM enabled. In fact I have
a fork
of gperftools containing some additional tcmalloc hacks I've found useful.
One problem with this of course is that much of cpython's memory is allocated
using its own memory pool allocator, which is largely invisible to the malloc
implementation. So I've also got
a patch for cpython
which adds a very basic canary mechanism to its pool allocator (at the slight expense of
memory efficiency).