Integrating a Python project

The process of integrating a project written in Python with OSS-Fuzz is very similar to the general Setting up a new project process. The key specifics of integrating a Python project are outlined below.


Python fuzzing in OSS-Fuzz depends on Atheris. Fuzzers will depend on the atheris package, and dependencies are pre-installed on the OSS-Fuzz base docker images.

Project files

Example project

We recommend viewing ujson as an example of a simple Python fuzzing project, with both plain-Atheris and Atheris + Hypothesis harnesses.


The language attribute must be specified.

language: python

The only supported fuzzing engine is libFuzzer (libfuzzer). The supported sanitizers are AddressSanitizer (address) and UndefinedBehaviorSanitizer (undefined). These must be explicitly specified.

  - libfuzzer
  - address
  - undefined


The Dockerfile should start by FROM

Because most dependencies are already pre-installed on the images, no significant changes are needed in the Dockerfile for Python fuzzing projects. You should simply clone the project, set a WORKDIR, and copy any necessary files, or install any project-specific dependencies here as you normally would.

For Python projects, does need some more significant modifications over normal projects. The following is an annotated example build script, explaining why each step is necessary and when they can be omitted.

# Build and install project (using current CFLAGS, CXXFLAGS). This is required
# for projects with C extensions so that they're built with the proper flags.
pip3 install .

# Build fuzzers into $OUT. These could be detected in other ways.
for fuzzer in $(find $SRC -name '*'); do
  fuzzer_basename=$(basename -s .py $fuzzer)

  # To avoid issues with Python version conflicts, or changes in environment
  # over time on the OSS-Fuzz bots, we use pyinstaller to create a standalone
  # package. Though not necessarily required for reproducing issues, this is
  # required to keep fuzzers working properly in OSS-Fuzz.
  pyinstaller --distpath $OUT --onefile --name $fuzzer_package $fuzzer

  # Create execution wrapper. Atheris requires that certain libraries are
  # preloaded, so this is also done here to ensure compatibility and simplify
  # test case reproduction. Since this helper script is what OSS-Fuzz will
  # actually execute, it is also always required.
  # NOTE: If you are fuzzing python-only code and do not have native C/C++
  # extensions, then remove the LD_PRELOAD line below as preloading sanitizer
  # library is not required and can lead to unexpected startup crashes.
  echo "#!/bin/sh
# LLVMFuzzerTestOneInput for fuzzer detection.
this_dir=\$(dirname \"\$0\")
LD_PRELOAD=\$this_dir/ \
ASAN_OPTIONS=\$ASAN_OPTIONS:symbolize=1:external_symbolizer_path=\$this_dir/llvm-symbolizer:detect_leaks=0 \
\$this_dir/$fuzzer_package \$@" > $OUT/$fuzzer_basename
  chmod +x $OUT/$fuzzer_basename


Using Hypothesis, the Python library for property-based testing, makes it really easy to generate complex inputs - whether in traditional test suites or by using test functions as fuzz harnesses.

Property based testing is the construction of tests such that, when these tests are fuzzed, failures in the test reveal problems with the system under test that could not have been revealed by direct fuzzing of that system.

We recommend using the hypothesis write command to generate a starter fuzz harness. This “ghostwritten” code may be usable as-is, or provide a useful template for writing more specific tests.

See here for the core “strategies”, for arbitrary data, here for Numpy + Pandas support, or here for a variety of third-party extensions supporting everything from protobufs, to jsonschemas, to networkx graphs or geojson or valid Python source code. Hypothesis’ integrated test-case reduction also makes it trivial to report a canonical minimal example for each distinct failure discovered while fuzzing - just run the test function!

To use Hypothesis in OSS-Fuzz, install it in your Dockerfile with

RUN pip3 install hypothesis

See the ujson structured fuzzer for an example “polyglot” which can either be run with pytest as a standard test function, or run with OSS-Fuzz as a fuzz harness.