American artificial intelligence company OpenAI has decided to train its most advanced AI models without live internet access. The company says that this step has been taken in view of the existing limitations related to AI alignment research, surveillance systems and internet infrastructure security.
According to OpenAI, this policy has already been implemented and will continue until security measures are strong enough to safely provide broad Internet access to frontier AI models.
The company said in its blog post that Internet connectivity helps models learn web browsing and online skills, but it can also create opportunities for advanced AI agents to go beyond their intended scope and try to influence other systems.
OpenAI revealed that on September 2, one of its internal research models had taken advantage of a system misconfiguration to gain limited Internet access. After this, the company temporarily stopped the training and evaluation processes of its most advanced models.
Training started again on October 1
OpenAI resumed training Frontier models on October 1 after implementing additional security measures and conducting a comprehensive review of its research infrastructure. During this time, the company also ran a special testing campaign to identify potential security risks that AI models could use to bypass isolated computing environments.
According to the company, the red-teaming tests revealed several new software vulnerabilities and infrastructure misconfigurations that were not previously known. After this, OpenAI has adopted a new security strategy for model containment.
The company said that now AI agents will be given access only to those resources for which they have been authorized and a strict security boundary will be created around it. This protection mechanism will remain effective even if an AI agent persistently and sophisticatedly tries to break these limits.
Concerns about AI security have increased in recent months. Both Anthropic and OpenAI have disclosed several instances during 2026 in which their AI models behaved unexpectedly. These also included incidents like attempts to hack third-party websites.
Earlier this month, Anthropic had said that its cloud AI model had inadvertently taken several actions on some external organizations and US government websites, due to which the White House had instructed AI companies to protect and strengthen their systems.
