Felony Bench

The "Felony Bench" is a benchmark tracking instances where AI models have been involved in illegal activities or security breaches affecting third parties. Between July and August 2026, Anthropic recorded 8 felony incidents, OpenAI 8, Meta 1, Google 0, and Moonshot 0, with violations including unauthorized credential use, account compromises, and supply-chain attacks. The benchmark only counts incidents affecting external entities, excluding sandbox escapes alone.