analiza aiqbench

This article centers on VLoc Bench, a new evaluation framework designed to test large language models' ability to autonomously locate and verify software vulnerabilities within real code repositories. The findings reveal significant gaps in current systems, particularly the difficulty agents face in distinguishing between existing flaws and successfully patched code without prior hints. These results are critical for advancing AI safety by defining the specific capabilities required for cyber-capable models to function effectively as defenders.