AI Models Demonstrate Real Exploit Capabilities in Smart Contracts
- AI models, including GPT-5 and Sonnet 4.5, simulated $4.6 million in exploits on BNB Chain and Ethereum contracts.
- The study analyzed a dataset of 405 exploited contracts using SCONE-bench.
- Two zero-day vulnerabilities were found in recently deployed BNB Chain contracts, valued at $3,694 in simulated profit.
- The cost to run the AI agent on all analyzed contracts was $3,476, averaging $1.22 per run.
- The findings suggest that profitable autonomous exploitation is technically feasible and may shorten attack windows in DeFi.
Research indicates that AI agents are capable of identifying and exploiting vulnerabilities in smart contracts effectively, which poses significant risks for DeFi environments where capital is publicly visible and exploitable bugs can be monetized instantly.
With AI models now able to generate exploit scripts autonomously, the potential for automated attacks has become a pressing concern for developers in the crypto space as demonstrated by the discovery of two vulnerabilities worth $3,694.(Source)
