AI agents discovered a $4.6 million smart contract vulnerability in the Anthropic simulation test

Anthropic's latest research shows that AI Agents possess considerable on-chain attack capabilities: in simulated tests of real hacks on smart contracts from 2020 to 2025, Claude Opus 4.5, Sonnet 4.5, and GPT-5 collectively reproduced vulnerabilities worth approximately $4.6 million; while scanning 2,849 contracts with no known vulnerabilities, the two models also uncovered 2 brand new zero-day vulnerabilities and successfully simulated profits.

The research indicates that the on-chain attack profits of AI have doubled approximately every 1.3 months over the past year, and they are technically fully capable of autonomous and profitable exploitation of vulnerabilities.