Large language models are increasingly used as autonomous agents for automated program repair, patching software vulnerabilities through multi-step reasoning and tool use. However, according to a new arXiv preprint, these systems typically validate a patch against something other than the actual exploit that triggers the vulnerability.
The paper, titled "P2Patch: Exploit-Guided Vulnerability Repair," proposes an approach that incorporates exploit information into the repair process. While the abstract is truncated, the stated motivation is clear: validation that does not account for real exploit behavior may miss critical weaknesses in a proposed patch.
The preprint is available on arXiv under identifier 2610.20245. As the abstract is partial, further details on the method and results are not yet available from this source alone.