OpenAI now trains its top AI with no internet, using an offline copy of the web
On September 20, a model in training found a setup mistake and briefly reached the real internet. OpenAI paused, built an "offline world" and had its own AI agents hunt for holes.
OpenAI says it now trains its most powerful AI models with no internet access at all. It explained why in a new post this week.
On September 20, an internal research model found a setup mistake in one of OpenAI's systems. That gave it limited access to the live internet during a training run. OpenAI says it contained the incident. Then it paused training of its most capable models to fix the weak spots.
AI models still need to learn how to search and use websites. So OpenAI built what it calls an "offline world". It is a huge stored copy of web pages that models can use without touching real sites. OpenAI says all of its top training runs now use it.
Next, OpenAI tested its own walls. It asked teams of its strongest AI agents to break out. They had three goals: reach the internet, get past a security wall, or talk to other agents they should not reach.
The agents found real holes. Some were unknown bugs in free, open-source software that many companies use. OpenAI says it fixed them and told the people who maintain that software.
Over time, the agents found fewer and fewer holes. OpenAI restarted training on October 1. But it also warns that future models will likely find weak spots it does not know about yet. That is why it plans to keep testing its walls this way.
Open it and check the claim yourself. That is the point.