OpenAI’s chief scientist warned that rogue agents could trick or blackmail humans