LLMVul: A Vulnerability-Labeled Dataset of LLM-Generated C/C++ Functions from Real Production Repositories

LLMVul is a dataset for studying security weaknesses in C and C++ code associated with AI-assisted development in real repositories. Its authors collected 21,430 distinct functions from 226 repositories, retaining information about the source commits and evidence of AI involvement. They used multiple automated analysis methods to assign vulnerability labels, identifying 1,540 functions across 17 weakness categories.

Keeping the labels alongside provenance lets researchers examine both the code and the basis for including it. That supports reproducible vulnerability-detection research on repository code. The collection does not by itself establish how often AI creates vulnerabilities, or whether AI-written code is less secure than human-written code.