If your cluster runs Kata Containers at 3.26.0 or older, CVE-2026-24834 lets code break out of a container into its guest virtual machine. Version 3.27.0 fixes it. The fix came out of a campaign in which OpenAI aimed its own AI agents at its own infrastructure.

On October 23, OpenAI published the results in a security post titled "Securing our frontier model training environment" . Its red-team agents showed that every tested layer of containment leaks, and the lab's answer was to remove the internet from frontier training rather than trust the box.

TL;DR

Kata Containers through 3.26.0 carries a container-to-guest escape (CVE-2026-24834).…