A Framework for Fuzz Target Generation and Evaluation

This framework generates fuzz targets for real-world C/C++/Java/Python projects with various Large Language Models (LLM) and benchmarks them via the OSS-Fuzz platform.

More details available in AI-Powered Fuzzing: Breaking the Bug Hunting Barrier:

Current supported models are:

• Vertex AI code-bison
• Vertex AI code-bison-32k
• Gemini Pro
• Gemini Ultra
• Gemini Experimental
• Gemini 1.5
• OpenAI GPT-3.5-turbo
• OpenAI GPT-4
• OpenAI GPT-4o
• OpenAI GPT-4o-mini
• OpenAI GPT-4-turbo
• OpenAI GPT-3.5-turbo (Azure)
• OpenAI GPT-4 (Azure)
• OpenAI GPT-4o (Azure)

Generated fuzz targets are evaluated with four metrics against the most up-to-date data from production environment:

• Compilability
• Runtime crashes
• Runtime coverage
• Runtime line coverage diff against existing human-written fuzz targets in OSS-Fuzz.

Here is a sample experiment result from 2024 Jan 31. The experiment included 1300+ benchmarks from 297 open-source projects.

Overall, this framework manages to successfully leverage LLMs to generate valid fuzz targets (which generate non-zero coverage increase) for 160 C/C++ projects. The maximum line coverage increase is 29% from the existing human-written targets.

Note that these reports are not public as they may contain undisclosed vulnerabilities.

Usage

Check our detailed usage guide for instructions on how to run this framework and generate reports based on the results.

Independent Agent Execution and Evaluation

You can also execute or evaluate individual agents without running full experiments, using the integrated agent execution framework. See the framework's documentation for detailed instructions on how to run individual agents or sequence of agents.

Collaborations

Interested in research or open-source community collaborations? Please feel free to create an issue or email us: [email protected].

Bugs Discovered