OpenAI has introduced EVMBench, a benchmark designed to assess AI agents' capabilities in understanding, repairing, and exploiting smart contracts. The benchmark utilizes a test set of 120 high-risk vulnerabilities sourced from 40 real-world projects. It focuses on three key tasks: vulnerability discovery, code repair, and attack simulation. This initiative is seen as a critical evaluation of AI agents' ability to operate autonomously and collaboratively in the crypto environment, addressing fundamental issues of AI's future role in blockchain ecosystems.