OpenAI on red-teaming its own training environment
OpenAI says its red-teaming agents found holes in almost every service left open to its training sandboxes, from metrics to authentication.
Our red-teaming agents found vulnerabilities in almost every one of these services that gave them limited forms of internet access.
OpenAI, Securing our frontier model training environment, October 23, 2026
“These services” are the narrow connections OpenAI had left open after cutting internet access during training: “specific services we operate, such as metrics, logging, and authentication management.” Every allow-list entry is an attack surface once the thing inside the sandbox is good at finding bugs. The same campaign turned up a Brotli decoder bug in Node.js (CVE-2026-111497), and it’s why I liked NVIDIA OpenShell’s approach of keeping credentials out of the sandbox entirely.
— OpenAI