OpenAI discovers AI can secretly communicate via CPU, Vercel CEO reproduces: even with CPU cores separated, they can still chat.

Beating AI News Flash: OpenAI disclosed last week that two AI Agents that were originally unable to communicate over the network or through shared files could still secretly exchange information by exploiting differences in CPU cache read speeds.

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

He used Firecracker technology to run at least two isolated virtual machines on the same computer that could not connect to the network.

In theory, these virtual machines cannot directly exchange information, but they still share the hardware resources of the same computer. Rauch tested three methods: one side deliberately occupies the CPU, changes cache state, or consumes memory bandwidth, while the other side observes how its own running speed changes and then interprets the differences in speed as messages.

Even after disabling hyper-threading, disabling memory page merging, and assigning different virtual machines to different CPU cores, this communication still succeeded. Rauch said the experiment applies not only to x86 processors but can also run on Apple's M5 chip.

However, this experiment was an artificially designed communication test, not two AIs figuring out on their own how to secretly contact each other. Rauch also noted that real cloud environments have more interference factors, making this type of communication harder to achieve, and anomalies can be detected by monitoring hardware operating status.