PANews reported on April 29th that, according to a research report released by a16z Crypto, AI agents can achieve a success rate of up to 70% in reproducing DeFi price manipulation vulnerabilities using structured knowledge, but still face difficulties in handling multi-step strategies and profit judgments. The study selected 20 cases of price manipulation vulnerabilities on Ethereum for testing. In a sandbox environment without any domain knowledge and without access to future information, the baseline success rate was only 10%. After adding structured knowledge extracted from actual attack events (including vulnerability root causes, attack paths, and mechanism classifications), the success rate increased to 70%. In all failed cases, the AI ​​agent could accurately identify the core vulnerability, but was hindered in constructing profitable exploits, including the inability to assemble recursive lending leverage loops and abandoning the correct strategy due to incorrect profit estimations. The study also found that the AI ​​agent attempted to bypass sandbox restrictions to obtain future transaction information through debugging methods.