According to CoinMeta, OpenAI disclosed last week that two systems, AI and agent, which originally could not communicate via the network or shared files, managed to exchange information by taking advantage of the different read speeds of CPU caching. The founder of Vercel and CEO Guillermo Rauch demonstrated a program called CPU-chat that ran two isolated virtual machines on the same computer. Three methods were tested: one party occupied CPU, changed the cache state, or consumed memory bandwidth, while the other party observed the changes in running speed and interpreted these differences as messages. Even when hyper-threading was disabled, memory page merging was turned off, and the virtual machines were allocated to different CPU cores, this communication was still successful. Rauch stated that the experiment was applicable to x86 processors and Apple’s M5 chips, but in a real cloud environment, there are more interfering factors, making communication more difficult to achieve, and abnormalities can be detected by monitoring hardware status.