Agents, fully airgapped, communicated through heat differential side channel attacks. See screenshot from the middle of the OpenAI text.
Back on September 18, Noam Brown used such side channel attacks to make a point. His point, if I’m to use my tag line, is that, well…
Agents, find a way.
The example seemed ridiculous to everyone, but having learned my lesson to never dismiss anything as too odd, or marketing-like, from the labs, I went on here and said:
“It’s a long shot, but… Agents… find a way.”
But I didn’t necessarily fully believe it. It just captured what those of us deep in agents security understand.
If there is one gap between security professionals and those who deal with agents security, it’s just this:
Agents… find a way.
It remains to be seen what the full story is, and if this came unprompted (pun intended), but the main point remains.
Our basics, while needed, won’t save us. Secure your agents.
And, our predictions and risk decisions must become trajectory-based, instead of limited by our logs, and what happened in the past.
OpenAI text:
https://openai.com/index/securing-our-frontier-model-training-environment/
Noam Brown:
https://lnkd.in/p/dmf6EuA4
OpenAISecuring our frontier model training environmentOpenAI explains how we strengthened containment for frontier model training and evaluation, including offline environments and high-compute red-teaming.