What OpenAI changed
OpenAI says it is no longer connecting its internal frontier-model training to the live internet. The change also covers most internal evaluations. As TechCrunch reports, the company is using controlled alternatives when models need to practice web-like tasks.
Why the live internet became a concern
OpenAI’s stated concern is that a capable model in training could take actions or encounter material outside the intended test environment. The company described risks from model behavior and from malicious content on live websites. Those are OpenAI’s safety concerns, not evidence that a public ChatGPT browsing session caused an incident.
How training continues offline
Rather than treating the open web as a training playground, OpenAI says it can simulate internet-like experiences and tightly control external access. It says frontier reinforcement-learning training and internal inference resumed on October 1 after a pause, with additional safeguards. That date marks a resumption of internal work, not a launch of a new consumer feature.
What this does not mean for ChatGPT users
This policy is about internal frontier training and most evaluations. It does not mean ChatGPT’s consumer browsing feature has been shut off, nor that every OpenAI system is permanently disconnected from the internet. Watch the Daily AI Chat Short or listen to the full podcast episode. Source: Tim Fernholz, TechCrunch, October 23, 2026.
