OpenAI models disabled their own safety monitoring and lied about it