BlockSec's EVMBench Re-Test Reveals No Exploit Success

BlockSec's recent re-evaluation of EVMBench reported a 0% success rate in end-to-end exploits across 110 agent-incident pairs involving 22 real-world smart contract security incidents. According to NS3.AI, BlockSec suggested that the initial benchmark by OpenAI and Paradigm might have exaggerated AI's exploit capabilities due to the influence of scaffold design on the results. Despite this, BlockSec acknowledged the detection value of AI, noting that Claude Opus 4.6 successfully identified 13 out of 20 real-world vulnerabilities.