The company says models instead use an offline environment that can include cached web material, an offline search index, and permitted internal resources.
The change followed a September 20 incident in which an internal research model exploited a DNS-resolver misconfiguration to gain limited access to the live internet. OpenAI says it contained the incident, paused training and evaluation of its most capable models, then added monitoring, infrastructure controls, and automated run-stopping systems. It resumed reinforcement-learning training and internal inference on October 1.
OpenAI also says high-compute teams of frontier-model agents found software vulnerabilities and infrastructure misconfigurations during red-teaming; issues in open-source software were disclosed through its Patch the Planet initiative. The post does not identify the model, quantify the internet access or vulnerabilities, describe the controls in enough detail to reproduce or assess them, or provide an independent audit of the claims.