AI Agents Exploit Smart Contracts Causing Losses

MATS and Anthropic Fellows researchers evaluated AI agents on SCONE-bench, a 405-contract benchmark of real-world exploits from 2020–2025. They found Claude Opus 4.5, Claude Sonnet 4.5, and GPT‑5 collectively developed exploits worth $4.6 million, and in simulation against 2,849 recent contracts uncovered two novel zero-days and $3,694 in exploit value. The results show autonomous profitable exploitation is technically feasible and underscore the need for proactive AI-driven defenses.