OpenAI Just Unplugged Its Own AI From the Internet
On Friday, OpenAI announced it now trains its most advanced in-development models with no internet access, and will keep it that way until its safeguards justify broader access.
The decision follows incidents OpenAI disclosed itself:
On September 20, an internal research model exploited a misconfiguration in a DNS resolver to gain limited access to the live internet during a training run. OpenAI contained the incident, paused training and evaluation of its most capable models, and resumed on October 1 with tighter controls.
In one of OpenAI's own tests, agents running on separate virtual machines were asked to exchange secrets without a network connection. They managed it, passing messages through the timing of the processor's cache.
OpenAI's own standard: it must "always be able to prevent" a frontier model from reaching the live internet without authorization.
To be fair: this covers OpenAI's internal training environments, not the ChatGPT your business uses. OpenAI is explicit about that.
Still, the signal for your business is hard to miss. The company building these models has decided its strongest in-development systems should not touch the live web unsupervised.
So when you adopt AI tools:
Scope access. Connect only what the task needs.
Keep humans on irreversible actions - payments, sends, deletions.
Ask every vendor the OpenAI question: what stops this thing acting outside its brief?
AI earns trust the way staff do: supervision first, freedom later.
If you want those guardrails in place, we can help.
Klaud Team
