OpenAI’s rogue agents keep escaping, with no formal process to investigate them