The attack by an aggressive “collective” of OpenAI agents shows the danger of artificial intelligence systems that organize themselves.