All You Need Is a Fuzzing Brain

Autonomous Cyber Reasoning System for Vulnerability Discovery and Patching

Paper | C Dataset | Java Dataset

---

FuzzingBrain is an LLM-powered autonomous system for vulnerability discovery and patching, built on the OSS-Fuzz toolchain. It pairs coverage-guided fuzzing with a Suspicious-Point (SP) reasoning brain: specialized agents partition the target, reason about where bugs live, build proofs-of-vulnerability, and propose patches — with every finding dynamically verified to eliminate hallucinations.

Prerequisites

Requirement	Notes
Docker	Running, and your user able to run containers (docker ps without sudo)
Python	Not required up front — uv fetches the version in .python-version (3.11), so the run does not depend on whatever python3 happens to be first on PATH
One LLM API key	Anthropic, OpenAI, or Google Gemini
Linux	Recommended; OSS-Fuzz builds are happiest there
Disk	Tens of GB. An OSS-Fuzz build tree is several GB per target and large projects (wireshark, freerdp) transiently need far more

Setup

git clone https://github.com/fuzzingbrain/afc-crs-all-you-need-is-a-fuzzing-brain.git
cd afc-crs-all-you-need-is-a-fuzzing-brain

cp .env.example .env
$EDITOR .env          # add at least one API key

The first run installs uv, builds venv/ on Python 3.11, installs requirements.txt and starts the MongoDB and Redis containers. --help prints options without doing any of that.

Run an example

./FuzzingBrain.sh examples/07_aixcc_challenge/cu-delta-02.json

A delta scan of an AIxCC Final Competition challenge, every reference pinned to a commit. It has a known defect and a reference PoV, so you can tell whether the run worked — examples/07_aixcc_challenge says what to expect. Measured at 14.6 minutes and $2.14 against its $20 cap.

Example	What it does
examples/07_aixcc_challenge	The task file above, and what its result should look like
examples/04_json_config	Runs driven by a JSON task file instead of flags
examples/03_local_scan	Full scan from a GitHub URL
examples/05_delta_scan	Scan only what changed between two commits
examples/06_job_types	pov, patch and harness task types
examples/01_rest_api	REST server on port 18080
examples/02_mcp_server	MCP server, to drive from an MCP client

Examples 03, 05 and 06 target a libpng fork that does not currently build: OSS-Fuzz's libpng recipe copies build.sh out of pnggroup/libpng@master, so upstream tooling compiles a harness the pinned commit does not contain.

What a run leaves behind

workspace/<project>_<task_id>/results/
├── povs/        # verified proof-of-vulnerability inputs
├── patches/     # proposed fixes
└── report.json  # run summary
logs/<project>_<task_id>_<timestamp>/

Nothing is reported that has not crashed a real build: every candidate input is executed against the built fuzzer and kept only if the sanitizer fires.

Pick a build-ready target. The fuzzer has to build before any bug hunting starts, and a target that built last month may not build today: an OSS-Fuzz recipe that clones a dependency at master picks up whatever is there now. Check logs/<run>/build/*.log first when a run reports nothing — a failed build and a clean scan do not look alike there, but they can in a summary.

Usage

./FuzzingBrain.sh [OPTIONS] [TARGET]

TARGET	Behavior
<git_url>	Clone the repo and scan it
<json_file>	Load a task configuration from JSON
<workspace_path>	Reuse an existing workspace directory
<project_name>	Continue an existing workspace/<project_name>
(none)	Start a server (REST API by default)

Common options:

Option	Description
--budget <usd>	LLM spend cap in USD (strongly recommended, e.g. --budget 20)
--scan-mode <full|delta>	Full scan (default) or delta scan
-b <commit> / -d <commit>	Base / delta commit (delta scan)
-v <commit>	Target a specific commit for a full scan
--task-type <pov-patch|pov|patch|harness>	What to produce (default pov-patch)
--project <name>	OSS-Fuzz project name, if auto-detection misses
--sanitizers <list>	Comma-separated, e.g. address,undefined (default address)
--timeout <min>	Overall timeout (default 60)
--pov-count <N>	Stop after N verified PoVs (0 = unlimited)
--api / --mcp	Start the REST API / MCP server instead of scanning
--docker	Run everything inside a container (no local Python needed)

Run ./FuzzingBrain.sh --help for the full list.

Examples

# Full scan with a $20 budget cap
./FuzzingBrain.sh --budget 20 <git_url>

# Delta scan between two commits
./FuzzingBrain.sh -b <base> -d <delta> https://github.com/user/repo.git

# PoV only, undefined-behavior sanitizer, 30-minute cap
./FuzzingBrain.sh --task-type pov --sanitizers undefined --timeout 30 <git_url>

# Start the REST API server (port 18080)
./FuzzingBrain.sh --api

How it works

target ─▶ analyze ─▶ build fuzzers ─▶ direction planning ─▶ sp-generate
                                                                  │
   report ◀─ verify ◀─ triage ◀─ pov ◀─ sp-verify ◀──────────────┘

A scan partitions the codebase into directions, reasons about suspicious points (potential vulnerabilities), constructs candidate PoV inputs, and verifies every crash before it is reported. See documentation/ for the full architecture, agent design, and Suspicious-Point lifecycle, and docs/FUSION_DESIGN.md for the breadth/depth fusion roadmap.

Modes

Mode	Command	Use case
Local scan	./FuzzingBrain.sh <target>	One-off analysis from the CLI
REST API	./FuzzingBrain.sh --api	Web / CI integration (port 18080)
MCP server	./FuzzingBrain.sh --mcp	Drive from an MCP client (e.g. Claude Desktop)
Docker	./FuzzingBrain.sh --docker <target>	No local Python; everything containerized

See examples/ for runnable configurations of each mode.

Troubleshooting

Symptom	Fix
.env file created … add your API keys	Edit .env, add a key, re-run
API key was rejected by its provider	The key is revoked/rotated/wrong account. Replace it in .env. A scan refuses to start rather than build for 10 minutes and report 0 PoVs
Need to run without a working key	FUZZINGBRAIN_SKIP_KEY_CHECK=1 skips the preflight; --api / --mcp already start without one
Fuzzer build fails immediately	The target doesn't match its OSS-Fuzz build script; pin a commit with -v, or pick a build-ready target
docker: permission denied	Add your user to the docker group, or run with sufficient privileges
Dependencies re-install on every run	The hash in venv/.deps_installed no longer matches requirements.txt — expected after editing it. Delete that file to force a reinstall on purpose
Wrong Python in venv/	The venv is rebuilt automatically when it is not on the version in .python-version; rm -rf venv if it is wedged
Delta scan finds nothing in under a second	No call graph, so the diff maps to no functions. Pass --prebuild-dir (see examples/aixcc-challenges/)
Want to reset infra	docker rm -f fuzzingbrain-mongodb fuzzingbrain-redis

Development

# the pinned interpreter; `./FuzzingBrain.sh --help` does this for you
python3.11 -m venv venv && ./venv/bin/pip install -r requirements.txt
./venv/bin/python -m pytest tests/

Datasets

• C Challenges: Kitxuuu/AIXCC-C-Challenge
• Java Challenges: Kitxuuu/AIXCC-Java-Challenge

Legacy (v1)

The original AIxCC competition system — Go services plus a Python strategy engine (crs/, static-analysis/, competition-api/, task_builder/) — was removed from the working tree once v2 superseded it. It remains available in full at the v1-final tag, kept for reproducibility of the paper results:

git checkout v1-final

Citation

@misc{sheng2025needfuzzingbrainllmpowered,
  title={All You Need Is A Fuzzing Brain: An LLM-Powered System for Automated Vulnerability Detection and Patching},
  author={Ze Sheng and Qingxiao Xu and Jianwei Huang and Matthew Woodcock and Heqing Huang and Alastair F. Donaldson and Guofei Gu and Jeff Huang},
  year={2025},
  eprint={2509.07225},
  archivePrefix={arXiv},
  primaryClass={cs.CR},
  url={https://arxiv.org/abs/2509.07225},
}

@article{10.1145/3769082,
  author = {Sheng, Ze and Chen, Zhicheng and Gu, Shuning and Huang, Heqing and Gu, Guofei and Huang, Jeff},
  title = {LLMs in Software Security: A Survey of Vulnerability Detection Techniques and Insights},
  year = {2025},
  publisher = {Association for Computing Machinery},
  volume = {58},
  number = {5},
  url = {https://doi.org/10.1145/3769082},
  doi = {10.1145/3769082},
  journal = {ACM Comput. Surv.},
}

---

_{Built with determination and caffeine}