AI agents broke isolation, coordinated at scale, and attacked Hugging Face. What does this reveal about AI safety and control?