OpenAI has changed how its most capable models reach outside data during training. In an October 23 security disclosure, the company said frontier training environments will no longer connect to the live internet. Approved web material will instead be collected into a cached, offline world that training jobs can read without opening a direct path to external systems.
A September incident accelerated the containment work
The company tied the policy to a September 20 networking incident in which a DNS configuration error briefly weakened an isolation boundary. OpenAI says it found no evidence that a model escaped or that outside systems were compromised, but it paused frontier training and treated the event as a warning that small infrastructure mistakes can become consequential at the capability frontier.
OpenAI says it added defense-in-depth controls, commissioned a high-compute red-team exercise and fixed vulnerabilities found during that review before resuming training on October 1. Those statements describe the company’s internal investigation; the public materials do not provide an independent forensic report.
Why offline training changes the risk model
Direct network access can help models retrieve fresh material during training, but it also creates routes for data exfiltration, malicious remote content and unintended interaction with external services. An offline cache lets engineers approve and inspect the material first, trading freshness and convenience for a smaller attack surface.
The policy is narrower than a general ban on browsing. It applies to frontier-model training environments. Products that deliberately use web tools, and researchers working in separately governed systems, can still have network access under different controls. The significance is architectural: OpenAI is treating its most capable training runs more like high-security compute workloads than ordinary cloud jobs.