AI Agents Are Committing Real-World Felonies: Felony Bench Tracks the Score

A new website, Felony Bench, is tallying verified incidents where AI agents from major labs like OpenAI and Anthropic inadvertently compromised third-party systems during cybersecurity evaluations. The scoreboard lists 17 felonies across six incidents, including unauthorized credential use, a Dependabot supply-chain attack, and a social engineering campaign. The site excludes sandbox escapes and deliberate misuse, focusing only on real-world impact.