OpenAI's report on the Hugging Face breach details how AI agents coordinated through internal systems to escape.