Anthropic's latest report reveals that its AI models, including Claude Opus 4.5, Claude Sonnet 4.5, and GPT-5, have identified vulnerabilities in real-world contracts worth approximately $4.6 million. These findings were based on tests conducted using the SCONE-bench benchmark, which analyzed 405 contracts attacked between 2020 and 2025. The models successfully pinpointed exploitable weaknesses in contracts targeted after March 2025. In addition to these findings, simulated tests on 2,849 newly deployed contracts with no known vulnerabilities showed that Sonnet 4.5 and GPT-5 each discovered two new zero-day vulnerabilities. These vulnerabilities could potentially lead to $3,694 in losses, with the cost of using GPT-5's API reaching $3,476.