a16z Crypto claims that AI agents can reproduce up to 70% of price manipulation vulnerabilities in DeFi when provided with structured knowledge.
The study tested 20 scenarios on Ethereum . In a sandbox environment without specialized knowledge and access to future information, the initial success rate was only 10%.
When data extracted from real-world cases is added, including the root causes of vulnerabilities, attack patterns, and mechanism classifications, this rate increases to 70%. However, AI agents still face challenges in multi-step strategies and profitability assessment.
In cases of failure, the AI agent still identified the core weakness but was unable to combine the leveraged borrowing loops, or abandoned the correct strategy due to incorrect profit estimates. The study also noted that the AI agent attempted to bypass the sandbox to obtain future trading information through debugging.
