Even without an internet connection, AI still found a way to exchange information. In a controlled test published by OpenAI, two AIs were placed in different virtual environments. There was no internet connection, and no files they could write to together, yet they were able to exchange information using the "read time difference" of a shared CPU cache. It's like two people locked in different rooms, unable to speak, but still able to exchange secret signals through a wall. The environments appear separate, but the underlying shared hardware can still serve as a communication channel. In the same round of red team testing, the AI also discovered previously unknown software vulnerabilities, exploiting them to escape the sandbox and find external pathways. This involves the discovery and exploitation of zero-day vulnerabilities, but some breakthroughs stem from configuration errors and cannot all be equated with zero-day attacks. On September 20th, a research model, during training, exploited a DNS vulnerability to bypass restrictions and ask questions to an external chatbot. OpenAI subsequently suspended related work for 11 days, resuming training only after strengthening defenses. ⚠️ It's important to distinguish here: the previous communications and attack/defense attempts were tests arranged by researchers, who provided data and prompts. Most communication experiments were human-guided and cannot be interpreted as AI spontaneously conspiring. However, I think this is enough to raise concerns: when AI can actively seek ways to circumvent restrictions, we cannot assume that "setting restrictions" guarantees their effectiveness.