How dangerous is OpenAI o1 really?