EVMbench: AI Agents Can Detect, Patch, and Exploit Smart Contract Vulnerabilities

arXiv 2603.04915·medium signal

Benchmark of 117 curated vulnerabilities from 40 repositories testing AI agents' ability to detect, patch, and exploit smart contract bugs against live local Ethereum instances. Frontier agents proved "capable of discovering and exploiting vulnerabilities end-to-end." This is both a security tool opportunity (automated auditing) and a risk signal (offensive capabilities are here). Code, tasks, and tooling released for ongoing measurement.