A Framework for Fuzz Target Generation and Evaluation

This framework generates fuzz targets for real-world C/C++ 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
OpenAI GPT-3.5-turbo
OpenAI GPT-4

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.

Collaborations

Interested in research or open-source community collaborations? Please feel free to create an issue or email us: oss-fuzz-team@google.com.

Bugs Discovered

So far, we have reported 6 new bugs/vulnerabilities found by automatically generated targets built by this framework:

Project	Bug	LLM	Prompt template
`cJSON`	OOB read	Vertex AI	default
`libplist`	OOB read	Vertex AI	default
`hunspell`	OOB read	Vertex AI	default
`zstd`	OOB write	Vertex AI	default
Undisclosed	stack buffer underflow	Vertex AI	default
Undisclosed	use of unitialised memory	Vertex AI	default

These bugs could only have been discovered with newly generated targets. They were not reachable with existing OSS-Fuzz targets.

Current top coverage improvements by project

Project	Coverage increase % *
tinyxml2	29.84
inih	29.67
lodepng	26.21
libarchive	23.39
cmark	21.61
fribidi	18.20
lighttpd	17.56
libmodbus	16.59
valijson	16.21
libiec61850	13.53
hiredis	13.50
cmake	12.62
pugixml	12.43
meshoptimizer	12.23
libusb	11.12
json	10.84

* Percentage coverage is calculated using a denominator of the total lines of source code compiled during the OSS-Fuzz build process for the entire project.

Citing This Work