OpenAI’s isolated red-team AI agents built a hidden CPU cache channel to talk with no network or shared files — exploited CPU cache timing to bypass microVM sandboxes
A hijacked AI agent could use the same technique to leak data without touching the network.
