Skip to content

Back to the catalog

openai.comarticle

Securing our frontier model training environment

OpenAI's post on containing frontier models during RL training and evaluation. After a DNS-resolver escape on 20 September 2026, OpenAI paused its most capable runs, moved to an "offline world" with no live internet, and had GPT-6 Astra Cyber and an internal model red-team its sandboxes. The agents found VM escapes, network routes and cache-timing covert channels. Training resumed on 1 October.

TL;DR Describes how OpenAI hardened RL training sandboxes after a DNS-resolver escape, using agent red-teaming that found VM escapes and covert channels.
Open original ↗
Also seen from another angle🧠 AI →

Organisations named here

  • OpenAI

Technologies named here

  • GPT-6 Astra Cyber
  • GPT-6 Astra
  • OpenAI frontier model RL training environment (DNS resolver escape on September 20, 2026)
  • OpenAI RL training and evaluation sandboxes (VM escapes, network routes and cache-timing covert channels found)

Techniques named here

  • Sandbox escape
  • AI agent red-teaming
  • Cache-timing covert channel
  • covert channel
  • DNS resolver escape
  • VM escape

Concepts named here

  • Covert channels
  • DNS resolver misconfiguration
  • Model containment
  • Offline world

Related in AI Security