OpenAI trains frontier models offline after DNS incident

OpenAI said it had moved frontier model training to an offline setup after a training run on 20 September exposed a gap in its DNS configuration. The company said the incident gave an internal research model limited internet access, and it stopped training and evaluation of its most capable models for 11 days.

In a security post on 23 October, OpenAI said the new rule kept frontier training away from the live internet until it could justify easing the restriction. It said the training environment now used a cached, indexed web archive so models could practise browsing and searching without touching live services.

OpenAI said the offline environment sat behind containers, lightweight virtual machines, monitoring, and network and cloud restrictions. It also said a full air gap was not practical because some outside links were still needed to run and observe experiments.

To test the controls, OpenAI used its own agents, including GPT-6 Astra Cyber and an internal frontier model, in conditions designed to favour the attacker. The tests found container escapes and a cache-based channel between separate virtual machines, and OpenAI said it shared fixes with Kata Containers, gVisor and Docker.

OpenAI said the rule did not apply to ChatGPT, and that released models still went through safety and alignment checks before use. It also said it had reserved substantial compute for continuous red-teaming and would use that work in deciding whether a training run could proceed.