PromptFuzz is an automated tool for generating fuzz drivers for libraries by mutating LLM prompts within a fuzz loop. It targets researchers and developers seeking to improve code coverage and discover security vulnerabilities in C/C++ libraries, offering a novel approach to fuzzing that leverages LLMs for API interaction discovery.

How It Works

PromptFuzz employs a feedback-driven fuzz loop that mutates LLM prompts to generate fuzz drivers. It extracts library context to construct these prompts, prioritizing mutations that explore complex API interrelationships guided by code coverage. The generated programs undergo rigorous sanitization for syntax, semantics, behavior, and coverage before being integrated with the LibFuzzer grey-box fuzzing engine.

Quick Start & Requirements

• Install: pip install promptfuzz
• Prerequisites: Python 3.8+, LLM API access (e.g., OpenAI API key for ChatGPT/GPT-4), LibFuzzer.
• Setup: Requires configuration of LLM API keys and potentially building target libraries with fuzzing instrumentation.
• Docs: https://github.com/FuzzAnything/PromptFuzz

Highlighted Details

• Achieved 40.12% branch coverage on tested libraries, outperforming OSS-Fuzz and Hopper.
• Detected 33 valid security bugs across multiple libraries, including CVEs in libaom, libvpx, and libTIFF.
• Supports multiple LLMs (Codex, Incoder, ChatGPT, GPT4) and prioritizes API combinations for mutation.
• Integrates with LibFuzzer for grey-box fuzzing.

Maintenance & Community

The project appears to be research-driven with contributions from academic institutions. Further community engagement channels are not explicitly listed in the README.

Licensing & Compatibility

The README does not specify a license. This lack of explicit licensing makes commercial use or integration into closed-source projects uncertain without further clarification.

Limitations & Caveats

PromptFuzz is currently tested on ChatGPT and primarily targets C/C++ libraries. Future work aims to generalize to binary programs and apply to closed-source libraries via LLM fine-tuning, indicating current limitations in these areas.