BlockSec researchers re-evaluated the EVMBench AI audit benchmark and found that its claimed 72% AI exploit success rate may be overly optimistic. In a paper titled "Re-evaluating EVMBench," they tested 26 agent configurations and analyzed 22 real security incidents that occurred after February 2026, with results showing a 0% end-to-end exploit success rate. While AI detection results for known vulnerability patterns were consistent with the initial research findings, BlockSec co-founder (Yajin Zhou) believes that fully automated auditing is not yet practical, advocating for a human-AI collaboration approach in which AI handles breadth analysis while humans handle depth analysis.