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

Comparatively, according to cutting-edge Red Team test results published by Anthropic, AI agents identified multiple smart contract vulnerabilities in a simulated blockchain environment, which could potentially be exploited up to $4.6 million.

The test covers complete processes such as contract analysis, command line tool configuration, network reconstruction, exploit generation and verification, and also introduces a new benchmark for smart contract security assessment. The study was jointly completed by Anthropic, MATS Project, and Fellows Program.