CoinWorld reports that OpenAI revealed last week that two AI agents, which originally had no way to communicate over a network or through shared files, successfully exchanged information by exploiting differences in CPU cache read speeds. Vercel founder and CEO Guillermo Rauch demonstrated a program called CPU-chat, running two isolated virtual machines on the same computer. Rauch tested three methods: one side occupied the CPU, changed the cache state, or consumed memory bandwidth, while the other observed changes in processing speed and interpreted the differences as messages. The communication still worked even with hyper-threading turned off, memory page merging disabled, and the virtual machines assigned to different CPU cores. Rauch said the experiment works on x86 processors and Apple’s M5 chip, but interference is greater in real cloud environments, making communication harder; anomalies can also be detected by monitoring hardware states.