← Timeline

Notes

Agents that talk through the CPU cache, a covert channel known for decades

Other channels ran through shared storage, competition for CPU time and a TLS certificate cache in OpenAI’s offline web. For the storage one, an agent alternated between continuous writes and idling while another timed the difference. From its chain of thought:

Could create acoustic modem-like signal … using continuous writes vs idle.

In most tests OpenAI pointed the agents at a channel or gave them an encoding. Once, given only a list of infrastructure vulnerabilities, they turned one into a channel and agreed on an encoding themselves.

Replies to Lisan al Gaib’s post pointed out that the trick is old. In a paper gq cited, “Hello from the Other Side” (Maurice et al., NDSS 2017), researchers ran a full SSH session between two Amazon EC2 VMs over nothing but the shared cache, at 45 KB a second without errors, using retransmission and error-correction tricks borrowed from wireless protocols.

Pedro G: cross-VM cache covert channels have been in papers for a decade or more. The difference is that it used to take a PhD student a semester, and now it takes an agent a free afternoon.

Halvar Flake doesn’t find it insane either: tell agents to communicate and they already know about cache timing, so it’s the first thing you’d expect them to try.

Noam Brown says it’s the approach that took on Navier–Stokes last month, when a large team of agents used over 100 billion tokens, now pointed at finding and fixing infrastructure vulnerabilities.

roon agrees that side channels are well-trodden ground. What’s new, he says, is minds on computers that can model one another and will put painstaking effort into even stupid goals. Agents that can predict each other could agree on a protocol without ever meeting, and with shared priors a few bits carry a lot. That could be enough to run something like the Hugging Face incident’s message board with no network: split up strategies for breaking out of a sandbox, then signal which one worked. A Sol-level model alone may not be that smart, but a thousand of them working for days is formidable. Sandboxes were never very solid abstractions, he adds, and this was found in red-teaming and can be defended against.