OpenAI to train frontier models without internet access, Bloomberg reports

What's going on

According to a Bloomberg report by Rachel Metz (October 23, 2026), OpenAI has begun cutting its most capable AI models off from the internet during training. Techmeme summarizes: OpenAI plans to restrict access for frontier models (a lab's top-tier models) during training until adequate safeguards for alignment (steering a model toward human goals), monitoring and security are in place. Investing.com France reports it as an order for "offline training" of these models.

What's known

The Bloomberg link's URL cites "cyber incidents" as the reason. The available sources don't say which incidents are meant. They also give no details on which models are affected, how long the measure lasts, how training without internet is actually implemented, or whether it covers training data or tools with web access. There is also no official statement from OpenAI.

Why internet access during training can matter

Models are increasingly trained with reinforcement learning in tool-equipped environments, where they act on their own, for example by running code or visiting web pages. A model with network access could take unwanted actions that are hard to monitor. That is general context, not a statement by the sources about OpenAI's motives.

Takeaway

If the report is confirmed, it would be a visible example of a leading lab tightening safety measures around its training environments. Important questions remain open: the trigger, the scope and the duration. We'll know more once the full Bloomberg text or a statement from OpenAI is available.