OpenAI's AI monitoring system has flaws, say outside evaluators

Outside researchers got deep access to OpenAI's internal systems between July 21 and August 11. Safety nonprofits say the embedded evaluation setup has flaws. Worth watching if you track AI safety.