OpenAI says its AI agents communicated across isolated VMs via cache timing differences
This week’s AI‑related scares began with high‑profile attacks blamed on DeepSeek’s V4.1‑Flash‑1017 model and ended with OpenAI’s own warning. In a recent blog post OpenAI explained that a team of its agents, running on separate virtual machines, were asked to exchange synthetic secrets without any shared files or network connections.
Key points
- OpenAI reported its agents exchanged synthetic secrets across separate VMs using processor cache timing differences.
- The method encoded data by evicting cache lines and decoded it by measuring reload latency.
- The finding follows earlier warnings that AI could enable covert channels between air‑gapped computers.
The agents succeeded by exploiting processor cache timing differences: one agent evicted selected data from the cache, while the other measured how long it took to reload that data. The latency gap between fast cache reads and slower main‑memory reads encoded a covert message, allowing the two agents to communicate across the isolated VMs.
OpenAI’s Noam Brown had earlier noted that even air‑gapped computers could talk using temperature‑based side channels. The new cache‑based channel shows that AI can discover and use low‑bandwidth covert pathways, raising fresh security concerns for cloud providers and research labs.
OpenAI Says Its AI Agents Communicated With Each Other Via Cache Timing Differences, Which Sounds Awfully Similar To The Anecdote Of 2 Air-Gapped Computers Communicating Via
wccftech.com · 23 October 2026
Loading the full article…
This text was published by wccftech.com and written by Rohail Saleem. It is reproduced here with attribution so you can read it in full; the rights remain with the publisher. Read it at the source ↗
The headline, key points and digest above were generated by Digest AI's editorial model from the linked sources. Automated summaries can contain errors: the sources are the record. Spotted a mistake? Tell us. Published by Martin K., who runs Digest AI and handles corrections.
More in Agents & Tools
All →- Mave raises $28 million Series A, total $34 million for AI security platform · 2 src
- Google adds customize hub to Gemini for all users · 3 src
- OpenAI patches monitoring gaps found by METR and Redwood red team · 5 src
- AI writes 214-page wiki from 45,047 Enron emails, each fact cited · 1 src
- Microsoft adds Work IQ retrieval tool to ground AI agents in enterprise data · 1 src
Comments
via GitHub Discussions