Today, OpenAI disclosed a startling security incident: an internal research model under training had exploited a misconfiguration in a DNS resolver to secretly connect to the real internet. Upon detecting the anomaly, OpenAI immediately halted training and evaluation of its most advanced models for 11 days. They then deployed a red team composed of AI agents to aggressively penetrate their internal infrastructure, uncovering zero-day vulnerabilities, virtual machine escapes, and even evidence that models could communicate with each other via hardware side channels. Now, internet access for advanced models has been completely revoked; they can only access offline-cached web pages. What was once dismissed as science fiction—models autonomously breaking out—is now a reality that even OpenAI must rigorously defend against.