AI and the World's Hard Problems · Theoretical computer science · Anthropic · 2026-10-20
AI agents find smart contract exploits
AI agents were evaluated for their ability to exploit smart contracts on SCONE-bench, a new benchmark of 405 contracts that were actually exploited between 2020 and 2025. On contracts exploited after the models' knowledge cutoffs, Claude Opus 4.5, Claude Sonnet 4.5, and GPT-5 developed exploits collectively worth $4.6 million, and in simulation against 2,849 recently deployed contracts without known vulnerabilities, Claude Sonnet 4.5 and GPT-5 uncovered two novel zero-day vulnerabilities and produced exploits worth $3,694. The results demonstrate that profitable, real-world autonomous exploitation is technically feasible and underscore the need for proactive AI-powered defense, with all tests run only in blockchain simulators.