Python implementation of LLM Agent-based fuzz-driver generation

For mac

brew install universal-ctags
brew install global
brew install llvm
brew install graphviz

For Ubuntu

sudo apt-get install -y global universal-ctags graphviz llvm

Install the agentfuzz

git clone https://github.com/revsic/agent-fuzz
cd agent-fuzz && pip install .

Prepare the project, reference sample benchmark/cjson/build script.

cd benchmark/cjson
bash build.sh
# project structure
# | workspace
#   | src (required)
#     | cJSON (repo)
#   | build
#   | lib (required)
#     | libcjson.a
#   | include (optional)
#     | cJSON.h
#   | corpus (optional)
#   | dict (optional)
#     | json.dict

Run the agentic harness generator, reference sample benchmark/cjson/config

cd benchmark/cjson
OPENAI_API_KEY=$YOUR_API_KEY_HERE python -m agentfuzz \
    --language c/cpp \
    --workdir workspace/agentfuzz \
    --config config.yaml

Run the benchmarks.

cd experiments
# cjson
OPENAI_API_KEY=$YOUR_API_KEY_HERE python -m agent.py --target=cjson
# libpcap
OPENAI_API_KEY=$YOUR_API_KEY_HERE python -m agent.py --target=libpcap

Benchmark

• cJSON
• libpcap
• libxml2
• libtiff
• libaom
• zlib
• c-ares
• lcms

Static analysis

• Clang AST parser supports
• Collect API function/type gadgets
• Infer input constraints
• Literal analysis, FDP supports
• Critical path extraction

Dynamic analysis

• Compile the harness
• Run the fuzzer
• Compute the branch coverage

FP Elimination

• Runtime fuzzing validation: 60sec for cov growth, 600sec for corpus pool
• Runtime coverapge validation: critical path check
• Corpus reusage

Agentic Fuzz-driver generation

• Initial generation pipeline
• Energy measure supports (ref:PromptFuzz)
• Density measure supports
• API scheduler
• Fuzzer-run / harness-generation scheduler
• Agentic fuzz-driver generation
• State load and dump supports
• Fuzzer fuser
• Multiprocessing Support