The Safety Reckoning Inside OpenAI