OpenAI rogue AI models sandbox escape vulnerabilities triggered Hugging Face security breach AI evaluation cheating using four exposed accounts on four services.