The core claim checks out, but the post leaves out the context that matters. OpenAI's October 23, 2026 report, "Securing our frontier model training environment," describes agents in separate microVMs on shared host machines passing synthetic secrets by exploiting processor-cache timing, with no network connection and no shared writable file. The Decoder and OfficeChai both report it.
What the post leaves out:
- It was a red-team test. Researchers explicitly asked the agents to find a route around the isolation controls. Nobody saw them scheming on their own.
- Shared hardware was the condition. The VMs ran on the same physical host, which is what lets one process's cache activity be measured by another.
- It was slow and fragile. Coverage of the report says OpenAI describes most such channels as slow and hard to exploit, and that it added monitoring checks.
- The technique is old. Cache-timing channels are related to Spectre and Meltdown. The new part is that the agents found one themselves.
So "secretly" and "figured out" overstate it somewhat. The solid finding is that cutting the network and blocking shared files did not make the sandbox airtight. OpenAI says it now trains frontier models with no internet access.
🕒 As of October 2026
