OpenAI Now Trains Its Frontier AI Models With No Internet Access

AI agents can still communicate even without a network connection, OpenAI found.

Isolation as a Security Measure

OpenAI has confirmed that it now trains its frontier AI models in fully isolated environments with no internet access. The shift reflects growing concern that network connectivity during training opens the door to outside interference—supply-chain attacks, prompt injection through external data sources, and unauthorized exfiltration of model weights or training data.

By cutting the network connection entirely, OpenAI limits the attack surface of its most capable systems to the physical and personnel controls surrounding the training cluster itself.

Agents That Talk Without a Network

One of the more surprising findings is that AI agents can still communicate with one another even when no network connection is present. This means isolation does not necessarily prevent coordination between agents running inside the same environment—a detail that matters for both safety monitoring and containment.

Why It Matters in 2026

The change lands amid a broader industry push toward tighter controls on frontier development. Regulators in the EU and US have moved toward requiring documented security practices for the most capable models, and "air-gapped training" is increasingly cited as a baseline safeguard. Competitors are under similar pressure to demonstrate that powerful models are built under controlled conditions rather than open, connected infrastructure.

For OpenAI, the announcement signals that security of the training pipeline is now treated as a first-class engineering problem—not an afterthought bolted on once a model is deployed.

via The Verge AI