EVMbench tests AI agents on smart contract security

Smart contracts hold over $100 billion in crypto assets, making their security paramount. Now, a new benchmark called EVMbench is putting AI agents to the test. Created with Paradigm, it evaluates whether AI can detect, patch, or exploit real-world vulnerabilities in Ethereum-style blockchain environments. The benchmark covers 117 high-severity bugs from 40 audits, including scenarios from the Tempo blockchain for stablecoin payments. AI agents are scored in three modes: finding vulnerabilities (detect), fixing them without breaking functionality (patch), and actually draining funds in a sandbox (exploit). The best model, GPT‑5.3‑Codex, scored 71% in exploit mode—a big leap over earlier attempts. To prevent cheating, exploits are run in an isolated environment and graded programmatically. Why it matters: As AI gets better at coding, it can either become a hacker’s weapon or a security guard—EVMbench helps steer it toward the latter.