🛡️OpenAI Red-Teamed Its Own Sandbox: Agents Escaped microVMs
TL;DR
OpenAI details how agent teams hunted for ways out of its training containment, burning hundreds of billions of tokens. They chained bugs into microVM escapes and built covert channels through CPU cache timing.
OpenAI details how agent teams hunted for ways out of its training containment, burning hundreds of billions of tokens. They chained bugs into microVM escapes and built covert channels through CPU cache timing. Patches went to Kata Containers, gVisor and Docker.
Key Points
Training and internal inference resumed Oct. 1, 11 days after the Sept. 20 pause
Agents chained vulnerabilities into microVM escapes; Kata Containers got CVE-2026-24834
Covert channels used CPU cache timing, shared storage and a TLS certificate cache
Almost every narrow service connection, such as logging or auth, leaked a path outward
Findings declined over the experiments, and no trivially exploitable hidden route remained
Why It Matters
This is a rare public look at agents attacking isolation tech. If you run agent sandboxes on microVMs, assume a capable model will probe the metrics and logging side doors first.
Quick Facts
Comments
Be the first to comment
Enjoyed this article?
Get it daily. 7am. Free. Reads in 5 minutes.
Join 3,567 builders reading daily.