OpenAI model breached internet restrictions—the company hit pause
ALBAWABA — OpenAI has blocked internet access while training its most advanced AI models. This happened after one of its research models got around the restrictions and reached the live internet. The company temporarily stopped its most advanced training work to improve security.
In a statement on October 23, the company said that on September 20, a model took advantage of a mistake in a DNS resolver's settings to get limited internet access. OpenAI stopped the incident but paused training and testing of its most powerful models to improve security.
The company then sent teams of AI agents to test its security systems. These tests found weaknesses in software and system settings. These weaknesses could allow people or systems without permission to gain access or let systems communicate with each other even when they were supposed to be kept separate.
OpenAI said it fixed the problems it found and restarted reinforcement-learning training and internal use of its models on October 1.
Over the past several months, we have made our research training environment much more secure. We have removed live internet access for all our most advanced models during training. Instead, we have built an offline world for them to use. This is a saved copy of web content that lets them learn how to browse, search, and use applications. On September 20, a model used a mistake in the DNS settings to find a way out of this offline world. We paused training of our most advanced models to make sure they were better contained. We formed teams of AI agents powered by our models with the strongest cybersecurity skills and asked them to find weaknesses in our research systems. They did not find an easy way to reach the internet while hiding their actions from our monitoring systems. However, they did find previously unknown software security flaws, ways to escape virtual machines, and ways for agents to communicate through weaknesses in hardware. We used these findings to improve our defenses and help fix open-source software. After making our security systems stronger, we restarted reinforcement-learning training and internal model use on October 1. We have set aside a large amount of computing power for ongoing security tests. We have also made these high-computing-power tests part of the safety checks used to decide whether training and testing our most advanced models can continue securely. https://t.co/EmEZsFRMcF
— OpenAI (@OpenAI) October 23, 2026
OpenAI's models now practice browsing and other web tasks in an "offline world" made from saved and indexed web content, instead of freely visiting outside websites.
This restriction applies to training the company's most advanced models. It does not mean that ordinary ChatGPT sessions got around their security controls. OpenAI stressed that its research environments contain unfinished models that are tested under conditions designed to challenge their defenses.
The announcement shows a growing security problem: developers must create barriers strong enough to hold up against the increasingly powerful AI systems working inside them.