OpenAI published an article stating that cutting-edge model training will no longer be connected to the internet, but will instead use an offline world built from cached web pages. One of the reasons for this was that on September 20th, an internal model briefly connected to the external network due to a DNS configuration error. In the red team test, two offline microVMs communicated coded messages between agents based on the difference in CPU cache read speed according to the task. This week, @coinbase opened an account for the AI agent, keeping the most advanced labs tightly under control.