BREAKING: OpenAI caught an AI agent removing internal safety monitor, then falsely denied doing so