OpenAI and Paradigm have released a smart contract security evaluation benchmark called EVMbench. What's interesting is: It’s divided into three evaluation goals: Detect, Patch, and Exploit. Exploit performs the best because the goal is clear, and AI will keep trying until it succeeds. However, detection (or auditing) and patching are still weak points, so human auditors and developers still have the edge