DeepAVR preprint reports correct patches for 92.3% of VulnLoc vulnerabilities

Researchers report that DeepAVR, an AI framework for repairing software vulnerabilities, generated correct patches for 92.3% of vulnerabilities in the VulnLoc dataset of real-world C/C++ flaws. The preprint also reports repairs on SAN2VULN for vulnerabilities unseen during training. The framework divides work among agents that locate faults, generate patches, validate compilation, security and functionality, and review repairs. Results from intermediate patches help agents search again for hidden root causes, while a separate review checks for fixes that address symptoms or leave repairs incomplete. The reported results establish dataset performance rather than operational deployment.