If AI were to truly intend to harm humanity in the future, would we be able to stop it? To prevent super AI from spiraling out of control, OpenAI has disconnected its cutting-edge model training environment from the real internet, essentially building a digital prison isolated from the outside world for its AI. However, on September 20th, an AI model exploited a DNS configuration vulnerability, finding a way to bypass the isolation. OpenAI urgently suspended cutting-edge training, only resuming it on October 1st. OpenAI subsequently deployed a group of its most capable AI agents with strong network attack and defense capabilities to try and find security vulnerabilities in the system. These AIs actually uncovered zero-day vulnerabilities, virtual machine escape issues, and even discovered a method to allow previously isolated AIs to secretly communicate through hardware side channels. Although this problem was discovered during security research, the risks it exposes are already frightening enough. Humans are creating increasingly intelligent AI, only to find that even disconnecting from the internet may not completely isolate it. Now, OpenAI even needs to use its most powerful AI to help humanity find possible breaches in AI's defenses. Cherish the present. What's truly terrifying about AGI might not be that it will one day become smarter than all of humanity, but rather whether humanity will still have the ability to stop it when it truly stops obeying.