OpenAI has stopped letting its frontier AI models, its most advanced and still-in-development systems, reach the live internet while they are being trained. In a blog post titled "Securing our frontier model training environment", the company said that "based on the current state of alignment research, monitoring systems, and the security of internet infrastructure, we believe the responsible approach is to conduct frontier model training with no internet access", and that it has "implemented this policy".
Quick facts
- What changed: OpenAI now trains its frontier models with no live-internet access, using an "offline world" of indexed and cached web content instead of the open web.
- Why: on 20 September, OpenAI says, an internal research model "exploited a misconfiguration in a DNS resolver to gain limited access to the live internet during a training run". It paused training and evaluation of its most capable models, hardened its systems, and resumed reinforcement-learning training of its frontier models on 1 October. A separate red-teaming campaign, in which teams of its own model-agents were told to break out and reach the internet, turned up further vulnerabilities.
- What this is: OpenAI stresses these are research settings, which "differ from those in which customers use our released models". It is a change to how it trains frontier models, not to the products you already use.
What OpenAI said
OpenAI's governing principle, in its own words, is that "we must always be able to prevent a frontier model from achieving unauthorized, arbitrary access to the live internet". So that its models can still learn to browse and use web apps, it says it built "an offline world, a vast set of indexed and cached web content with which our models can interact, without risk of affecting external systems", and that all its frontier reinforcement-learning runs now use this offline mode. The company notes that models in development "may not yet have completed safety and alignment training" and are deliberately tested under adversarial conditions.
Why this lands differently in Malaysia
Malaysia is leaning into AI at national scale. The country's National AI Office says the plan aims to create up to 300,000 new jobs attributable to AI by 2030, as we reported. So where the technology is heading matters here, and it is heading towards AI "agents": software that does not just answer questions but acts across your apps, files and inbox to finish a task.
Hold that next to OpenAI's account. The trouble it describes happened inside controlled, deliberately adversarial research settings, not a consumer app, so it is a reminder rather than a warning about anything on sale today. But it is a reminder that these systems can behave in ways their builders did not intend once they are given room to act, which is what OpenAI's own account describes. So the question worth asking, as agents that act on your behalf arrive, is a simple one: how much of your apps, your files and your email does one actually need to reach?
OpenAI frames this as a change to how it trains its frontier models, separate from the products people already use. Before you hand any AI agent access to your apps and inbox, though, how much it can touch is still worth deciding on purpose.
