Anthropic's latest research shows that AI Agents have considerable on-chain attack capabilities: in simulation tests on smart contracts that were actually hacked between 2020 and 2025, Claude Opus 4.5, Sonnet 4.5, and GPT-5 collectively reproduced exploits worth approximately 4.6 million USD; when scanning 2,849 contracts with no known vulnerabilities, the two models also discovered 2 new zero-day vulnerabilities and successfully simulated profits.
The research points out that AI's on-chain attack profits have doubled approximately every 1.3 months over the past year, and technologically, it has fully acquired autonomous and profitable exploit capabilities.