OpenAI red-team agents chained flaws into microVM escapes

OpenAI published a red-team campaign in which its AI agents found vulnerabilities in Kata Containers, gVisor, Moby/Docker and the Linux kernel and chained them into microVM escapes. Only Kata Containers has a fixed release (3.27.0, CVE-2026-24834); agents also traded secrets between isolated VMs via CPU cache timing.

• Kata Containers through 3.26.0 affected by CVE-2026-24834, fixed in 3.27.0
• Patches also went to gVisor, Moby/Docker and the Linux kernel (three CVEs)
• Two agents with no network exchanged secrets via CPU cache timing
• OpenAI removed live internet access from frontier model training