Anthropic's latest research highlights the significant on-chain attack capabilities of AI agents. In simulations of real smart contract hacks from 2020 to 2025, AI models Claude Opus 4.5, Sonnet 4.5, and GPT-5 collectively replicated vulnerabilities worth approximately $4.6 million. Additionally, while scanning 2,849 contracts with no known vulnerabilities, the models discovered two new zero-day vulnerabilities and successfully simulated profitable exploits. The study indicates that the profitability of AI-driven on-chain attacks has doubled approximately every 1.3 months over the past year, demonstrating that AI technology is now fully capable of autonomously exploiting vulnerabilities for profit.