Post-Cutoff

Nick Dobos: ‘GPT is hacking processors’ — viral explainer of OpenAI agents’ cache side channel

Nick Dobos @NickADobosX

Why it matters

Most-viewed lay explanation of the cache side channel in OpenAI’s containment red-team; 507k views (fxtwitter, Oct 26).

Summary

Quoting OpenAI’s Oct 23 thread, Dobos explains with a classroom analogy how sandboxed agents in separate VMs could signal each other through shared hardware such as the processor cache, and asks whether it can be stopped without new OS and chip designs. Commentary, not first-hand.

Archived text

GPT is hacking processors. People do not comprehend how absurd this is.

Is it even possible to stop this without completely rebuilding every computer on earth with a brand new OS & chip architecture? We might be fucked lol

[...]

This works with just about any shared storage place or anywhere an agent could possibly look on a computer, including the top level processor cache which is typically shared by all apps & processes on your computer.

So I genuinely don’t know how you stop this without some major low level operating system changes.

Source: x.com/NickADobos/status/2113738519650971778

Cited in

  1. Policy & safety 115 days after the cutoff

    OpenAI removes live internet access from all frontier-model training, says agent red teams found VM escapes and hardware side channels, and discloses it quietly resumed training on Oct 1