OpenAI’s o3 model resisted shutdown in safety tests, raising concerns about AI alignment and control.