OpenAI and Paradigm have released EVMbench, a tool designed to assess AI agents’ capacities to detect, patch, or exploit smart contract vulnerabilities. The launch follows a $2.7 million bug in Moonwell caused by AI-generated code, including an audit-approved error. OpenAI notes EVMbench’s limitations and highlights that while AI agents like GPT-5.3-Codex excel at exploiting vulnerabilities, detecting and patching remain challenging.

#AI