Security firm BlockSec retested EVMBench and concluded that the benchmark overestimates the automation capabilities of AI in smart contract auditing. They expanded the testing to 26 model configurations and introduced 22 real-world attack events occurring after February 2026. The results showed that in 110 tests, the AI achieved a 0% success rate in real-world exploitation, but its vulnerability detection performance was close to the original report, with some models able to identify known pattern vulnerabilities. (The Block)