EVMBench puts data behind something the past week already demonstrated on mainnet. Across 120 curated vulnerabilities, the best agents execute end-to-end exploits against ~72% of vulnerable contracts. But comprehensive detection (finding all high-severity issues across a full codebase) tops out around 46%. The bottleneck isn't security reasoning. It's search and coverage across large repositories. Meanwhile, AI-authored code is reaching production with the same classes of flaws auditors have been catching for years. Oracle miscalculations, unit conversion errors, broken scaling logic: introduced by AI, shipped without adequate review. Two different problems. Same underlying gap: security workflows haven't adapted to AI in development. This is the problem we built our security practice around.(Nethermind)