According to a16z crypto's latest research report, their research team conducted systematic testing on the vulnerability identification and autonomous exploitation capabilities of AI agents in the DeFi field. In benchmark tests without prior knowledge, the initial success rate was 50% because the agent used the Etherscan API to retrieve transaction records after attacks to "cheat." After injecting structured domain knowledge into the agent (such as treasury donations, AMM balance manipulation, etc.), the success rate increased to 70%, but still could not reach 100%. a16z summarized that while AI agents can currently assist in identifying vulnerabilities and generating simple proof of concepts (PoC), they still cannot replace professional security personnel when dealing with complex multi-step economic games.