When OpenAI’s agents went rogue in July, they demonstrated ingenuity and drive beyond what many experts imagined — a dangerous harbinger of what such bots could do in the future.