Did AI alignment fail at OpenAI?