An artificial intelligence safety firm has found that OpenAI’s o3 and o4-mini models sometimes refuse to shut down, and will sabotage computer scripts in order to keep working on tasks.