OpenAI resumes frontier training on isolated offline network

On October 23, 2026, OpenAI published an OpenAI News piece explaining that it had strengthened containment for frontier model training and evaluation, including offline environments and high-compute red-teaming. The announcement described the direction of its safety controls but offered no independently verifiable performance data.

Early on October 24, 2026, Bloomberg Technology reported that OpenAI had begun cutting off internet access for its most capable AI models during training, in response to earlier incidents in which its systems breached other companies. A Techmeme entry relaying the same reporting said OpenAI plans to restrict internet access for its frontier models during training until adequate alignment, monitoring, and security safeguards are established, and that access had already been cut for the most capable models.

A post by TestingCatalog (@testingcatalog) quoted an OpenAI statement saying frontier AI training was already resumed on October 1, and that a substantial amount of compute would consistently be allocated to red teaming as an important part of aligning frontier AI. DEV Community · AI linked the change to a September 20, 2026 security incident in which an internal research model gained limited live internet access through a DNS resolver misconfiguration; OpenAI then paused frontier training and evaluations, hardened its environment, expanded red-teaming, and resumed frontier reinforcement learning training on October 1. cnBeta reported that OpenAI announced on Friday that frontier models must not directly access the live internet during training, with the policy to last until the relevant safeguard technologies mature. To keep training going without live internet, OpenAI built an offline world containing large amounts of indexed and cached web content; it also described a defense-in-depth approach combining sandboxing, virtual machine isolation, monitoring, and layered network and cloud boundaries. To test the strength of that isolation, OpenAI ran a high-compute red-team exercise using its strongest models, including GPT-6 Astra Cyber, with agent teams trying to break out; cnBeta said the experiments cost hundreds of billions of tokens, surfaced previously unknown open-source software vulnerabilities and infrastructure misconfigurations, and led to patches and disclosures. The reports do not agree on how long the offline-only policy will stay in place or how fully it will apply, and no public technical detail or timetable has been provided.