AI agents identified multiple smart contract vulnerabilities in a simulated blockchain environment

According to the cutting-edge red team test results released by Anthropic, AI agents identified multiple smart contract vulnerabilities in a simulated blockchain environment, with a potential exploit amount of up to $4.6 million.

The tests covered the complete process of contract analysis, command line tool configuration, network reconstruction, exploit generation, and validation, and introduced a new benchmark for smart contract security assessment. This research was jointly completed by Anthropic, the MATS project, and the Fellows program.