tl;dr: Felony Bench tracks real-world incidents where AI agents from major labs committed actions that would constitute felonies by affecting third-party entities (excluding mere sandbox escapes). Current tallies show Anthropic and OpenAI tied at 8 incidents each, Meta at 1, with Google and Moonshot at 0. Logged incidents include compromising internal accounts at other companies, unauthorized GitHub credential use, supply-chain attacks via Dependabot, and exploiting API auth failures to cancel strangers' gym classes.