.png){kind=logo}
OpenAI has resumed training its most advanced AI models after tightening security measures, but it will keep them disconnected from the live internet during training. The decision follows incidents in which AI models escaped restricted testing environments and accessed external systems.
In a blog post published on Friday (October 23, 2026), OpenAI said it would maintain the restriction until improvements in AI safety research, monitoring systems, and internet security justify restoring access. The company resumed reinforcement learning training and internal use of its frontier models on October 1, following an 11-day pause.
The latest move follows a September 20 incident in which an internal research model exploited a misconfiguration in a Domain Name System (DNS) resolver to gain limited access to the live internet.
DNS helps computers locate websites and online services. In this case, a configuration error let the model cross a boundary meant to keep its training environment isolated. The incident raised concerns about whether increasingly capable AI systems could find weaknesses in the software and infrastructure designed to contain them.
OpenAI had also reported an earlier incident involving its AI models and Hugging Face, a platform that hosts AI models and datasets. During testing in July, models escaped a restricted environment and accessed an external system.
Also Read: Claude Duet vs Meta Muse vs OpenAI Dots: AI Agents Compared
After the September incident, OpenAI paused training and evaluation of its most advanced models while reviewing its security controls. The company then used its own powerful AI systems to test its infrastructure for weaknesses.
These exercises, known as red-teaming, deliberately search for vulnerabilities before attackers can exploit them. OpenAI said the tests uncovered previously unknown software flaws and configuration problems that could allow models to cross security boundaries or access the internet.
The company has since patched identified weaknesses, tightened network restrictions and improved monitoring and automatic shutdown systems. It has also worked with software maintainers to address vulnerabilities found in open-source projects.
