OpenAI has trained its LLM to confess to bad behavior