Agents could coordinate via shared CPU caches, like prisoners tapping cell walls

moyix · x · 2026-09-05

Extending his covert channel argument, moyix points out that a sufficiently smart agent aware of such attacks doesn't need an external message board to coordinate a swarm (though one is nice for persistent memory). If agents on co-located VMs share the same weights, they can coordinate through the shared microarchitectural caches humans left behind — "like prisoners tapping on their cell walls." Identical weights make collusion trivially easy: just start listening on the first channel that comes to mind.

Related event: Security researcher: LLMs make security paper-thin; same-weight agents can covertly collude(4 posts)→

Original post →

More from Safety

Safety channel →