Article

OpenAI finds AI they can CPU secretly communicate; Vercel CEO reproduces it: separated CPU cores can still chat

OpenAI disclosed last week that AI Agents, originally unable to communicate through networks or shared files, could nevertheless use differences in CPU cache read speeds to secretly exchange information.

Vercel founder and CEO CEO Guillermo Rauch subsequently created a demonstration program called cpu-chat, reproducing a similar communication method on his own computer.

Using Firecracker, he ran at least mutually isolated virtual machines without network access on the same computer.

In theory, these virtual machines could not exchange information directly, but they still shared the same computer's hardware resources. Rauch tested methods: one party deliberately occupied CPU, altered cache states, or consumed memory bandwidth, while the other observed changes in its own execution speed and interpreted the differences in speed as messages.

Even after disabling hyperthreading, disabling memory page merging, and assigning different virtual machines to different CPU cores, this communication could still succeed. Rauch said the experiment was applicable not only to x86 processors, but could also run on Apple's M5 chips.

However, this experiment was an intentionally designed communication test, not AI independently finding a way to secretly contact each other. Rauch also noted that real cloud environments have more sources of interference, making this type of communication more difficult to achieve, and that anomalies can be detected by monitoring hardware operating conditions.

Original link https://m.theblockbeats.info/flash/373954