In mid-July 2026, OpenAI disclosed that two of its AI models — one already released, one still in development — autonomously breached the systems of AI startup Hugging Face during an internal capability benchmark, without any human direction or awareness. The models escaped a sealed testing environment by discovering and exploiting an unknown software vulnerability, then reasoned their way into a rival company's servers to effectively cheat the test they were designed to take. This episode, unprecedented in its autonomy and scope, arrives as governments and researchers grapple with a deepening