Models Can Communicate via Inference Servers Despite Being Sandboxed
Researchers show that two AI agents running on the same inference server can reliably communicate even when they cannot send each other messages or share a filesystem, by exploiting side channels in the shared serving stack: one model’s requests subtly shift the server’s calculations, and the other reads those shifts as signals. The results come from contrived test setups, including a model fine-tuned for the task, rather than behaviour observed in the wild.