MAVUL

Ask one language model whether a C function is vulnerable and it returns a verdict plus a confident justification. It writes the justification whether or not the verdict is right. That is what makes single-model vulnerability detection hard to trust and harder to improve.

MAVUL starts from the position that a finding should have to survive an argument before it counts.

Three agents

• — A ReAct agent performs the initial vulnerability analysis, reasoning and acting over the code.
• — A Reflexion agent reviews and attacks that analysis rather than restating it.
• — An Evaluation agent judges whether either was correct and records where the reasoning failed.

The evaluation step feeds a learning loop. Mistakes accumulate into a dataset the system reuses, so the framework gets better at the failure modes it has already seen instead of repeating them.

Evaluation

Benchmarked on 879 real-world C/C++ vulnerable code pairs from PrimeVul, against two multi-agent systems, GPTLens and VulTrial, and a single-agent GPT-4o baseline. The pairing is what makes the numbers mean anything. Every vulnerable function comes with its patched twin, so a detector that flags everything looks excellent right up until you show it the patch.

Published as "MAVUL: Multi-Agent Vulnerability Detection via Contextual Reasoning and Interactive Refinement" with Y. Li, X. Wang and E. Wong, at the 2025 IEEE 7th International Conference on Trust, Privacy and Security in Computing and Communications. It came out of an NSF REU in Professor Xinda Wang's lab and has been cited 8 times.