trailofbits-buttercup/fuzzer/README.md

Demo fuzzer.

General idea here is we have a set of builders and a fuzzer bot managed through redis queues.

All commands are suggestions based on vibes at the moment and not really tested...

Building:

... have a linux box Install poetry

curl -sSL https://install.python-poetry.org | python3 -

pyenv install 3.8
pyenv global 3.8
poetry shell
poetry install

Demo prereqs

Nginx depedencies (this is to run the fuzz bot locally instead of in docker):

sudo apt-get install pcre2

git clone https://github.com/google/oss-fuzz.git
export OSS_FUZZ_PATH=$(pwd)/oss-fuzz

Running:

Start redis:

docker pull redis
docker run redis
export REDIS_IP=$(docker inspect \
  -f '{{range.NetworkSettings.Networks}}{{.IPAddress}}{{end}}' <container>)
export REDIS_URL="redis://$REDIS_IP"

Starting the orchestrator:

This orchestrator mocks the interactions of an actual orchestrator by receiving build outputs and adding them to a fuzzer distribution.

python fuzzing_infra/orchestrator.py --redis_url $REDIS_URL

Starting a build bot

This starts a build bot that will build fuzzr harnesses using helper.py. Allow caching means it will be built directly in the ossfuzz directory. This mode allows for only building a harness once rather than snapshotting per request trees.

python fuzzing_infra/builder_bot.py  --redis_url $REDIS_URL --allow-caching

Starting the fuzzer bot

python fuzzing_infra/fuzzer_bot.py --redis_url $REDIS_URL --timeout 120

Sending a build request:

Creating a manual build request:

python fuzzing_infra/stimulate_build_bot.py --redis_url $REDIS_URL --target_package nginx --ossfuzz $OSS_FUZZ_PATH --engine libfuzzer --sanitizer address

This command should result in the builder emitting build logs and then fuzzer logs.

On timeouts and OOMs

OSS Fuzz documents their settings for handling timeouts (anything which takes >65 seconds to reproduce) and out-of-memory (anything which uses >2.5GB of memory) here and here.

Timeouts

• Sometimes tracer-bot can hang. AIxCC has introduced a timeout flag to kill helper.py in v1.2.0.
• We set this timeout value to whatever the fuzzing harness's settings are, or default to 120 seconds (longer than oss-fuzz documents).

OOMs

• We set fuzzer-bot and tracer-bot to have at least 4Gi of memory during minikube testing.
• We acknowledge that a fuzzing harness can specify -rss_limit_mb=0 to disable memory limits, but these cases will likely be few and far between (upon examining the projects under oss-fuzz), and if it does get killed due to memory limitations, then our stacktrace will be empty and we can simply ignore and move on.