November 2, 2026 · Edition #39 · by Asaf Nakash

The wall they shared

I picture two offices with closed doors and disconnected phones. Someone can still tap a message through their shared wall. Separating the rooms hasn't removed everything they share. That is why I want to revisit OpenAI's earlier containment tests. Agents in separate microVMs, small virtual computers on shared hosts, were assigned to exchange a synthetic secret without network connections or shared writable files. They communicated through the processor's cache. The cache holds data for fast access. One agent deliberately pushed selected data out; the other measured whether reloading it was fast or slow. Those timing differences became a message under an agreed protocol. What looked like the computer doing its work was also one job signaling another. These were assisted, controlled tests using an established technique; OpenAI also reports communication through storage being busy or idle. What unsettles me is the gap between separation as we draw it and separation as an attacker can test it. Security plans describe intended interfaces: where messages go, who can open a file. Here, shared infrastructure supplied another route. I see creativity in that change of purpose: a route can be established research and still be missing from a defender's plan. I want red teams to challenge the belief underneath the controls. Give a system the objective of communicating across a supposed separation, with clear limits, and let the exercise explore possible routes rather than only a predetermined attack menu. A successful workaround can reveal a purpose for the infrastructure that its security design never considered. Faster hackers put pressure on how quickly security teams respond. Unfamiliar attack paths challenge what they believed couldn't happen. I want preparation for both, instead of assuming AI will bring only a quicker version of the attacker we already know.

Written by Asaf Nakash, Principal Product Manager for AI Security at Microsoft Defender and host of the Context Window podcast. Originally published in Context Window Edition #39, November 2, 2026.