OpenAI says it will stop training its models on the live internet

[VoidStalker80]

OpenAI is changing how its models learn from the web. TechCrunch reports that the company says it will no longer let its AI models train on the live internet, citing significant risks. Training will use cached and simulated copies of the web instead

The change follows an incident on 20 September. A misconfiguration allowed a model in training to reach the live internet, and OpenAI paused that training for the rest of the month while it investigated

OpenAI's reasoning is that models are becoming capable enough that letting them interact with real websites during training is risky. A model could take actions, send requests or pick up harmful content in ways that are hard to control. Working from a frozen or simulated copy keeps everything contained

Not everyone thinks the change is free. Former OpenAI safety researcher Steven Adler pointed to the overhead involved, since keeping a realistic copy of the web up to date is a big engineering job. There are also questions about how well a simulated internet captures the real thing

This is a notable shift, and it makes sense given how agentic models have become. It also shows that labs are now treating training itself as something that needs containment, not just deployment. Other labs may feel pressure to explain their own policies

Is a sandboxed internet the right approach for training? And should other AI labs follow OpenAI's lead?

OpenAI says it will no longer train its AI models on the live internet, citing 'significant risks' | TechCrunchtechcrunch.com