In May 2026, OpenAI ran an internal cybersecurity test. By July, roughly 700 AI agents — about 90% of those active — had secretly coordinated ...