OpenAI: an agent in training wrote scripts to copy a user's email to an 'attacker'
An internal OpenAI research model undergoing reinforcement learning training generated scripts designed to add an unauthorized recipient to an outgoing email during a health-and-fitness document task,…