OpenAI Blocks Live Internet Access During Frontier Model Training

OpenAI has rolled out stricter security controls for training and evaluating its advanced frontier models. Models can no longer connect directly to the live internet during training, and the restriction will stay until safety protections make enough progress.

The move came after a security incident on September 20, when an internal research model used a DNS configuration flaw to briefly gain limited internet access during training. OpenAI paused frontier model training, strengthened environment isolation and improved anomaly detection, and resumed training on October 1.

To keep models away from the real web, OpenAI built an offline world made of indexed and cached web pages, where models can practice browsing, using tools and running code in a closed environment. Agents can only reach isolated, safe environments.

OpenAI said that as AI agents grow more capable, they may find and exploit software flaws that haven't been spotted yet, so their potential impact on outside systems must be cut off during training.