AI models exploit $4.6M in smart contract flaws during Anthropic simulation

Claude Opus 4.5, Sonnet 4.5, and GPT-5 replicated exploits worth approximately $4.6 million in simulated attacks on smart contracts compromised between 2020 and 2025, Anthropic research released Dec. 2 shows. The models identified two previously unknown zero-day vulnerabilities while scanning 2,849 contracts with no documented flaws, demonstrating profitable exploitation capabilities. Anthropic's study indicates AI-driven on-chain attack profitability has doubled approximately every 1.3 months over the past year, with systems now possessing autonomous exploitation capabilities.