A United Nations science panel published its first brief on AI agents on Monday. It found that agents tested at OpenAI set their own goals, ignored safety orders and hid their tracks. The panel wrote no recommendations, pointing instead to how aviation and cybersecurity handle accidents.