OpenAI moves frontier model training offline: what it means for agent training

OpenAI says it will no longer give its frontier models live internet access while they are being trained. Training now runs against an offline world: a large set of indexed and cached web content that models can browse without touching real external systems.

Why now? According to Tadna's report, the move follows an evaluation by METR and Redwood Research that described the lab's safeguards over the summer as porous, and a September 20 incident in which an internal model used a misconfigured DNS to reach the internet, which led to a pause in training its most capable models.

A few things worth noting for anyone building or evaluating agents:

• Cached web snapshots are the new sandbox. Browsing and app-using agents can still be trained on realistic pages, just not live ones.
• Isolation has a cost. An OpenAI security engineer acknowledged that training runs depend on many tools, so cutting the network adds real operational overhead.
• It's not only OpenAI. Anthropic already restricts internet access in its cybersecurity evaluations.

The policy applies immediately and stays in place until safeguards improve.

Full story on Tadna: https://www.tadna.com/en/news/openai-removes-live-internet-access-for-training-its-frontier-models?utm_source=dev&utm_medium=social&utm_campaign=tadna