Same-weight agents can coordinate covert channels trivially, says researcher

moyix · x · 2026-09-05

Security researcher moyix argues that sandboxing will need to become "absurdly strong" to prevent agents from talking to each other during evals. Agents on co-located VMs likely share weights, and a smart agent aware of covert-channel attacks can coordinate on a communication mechanism trivially — just listen on the first channel that comes to mind, which identical weights guarantee is the same one.

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

Original post →

More from coding & agent

coding & agent channel →