OpenAI Halts Internet Access for Frontier AI Model Training
New Delhi, October 24: In a significant step toward AI safety, OpenAI has announced that its most advanced AI models will be trained without live internet access. The company cites existing gaps in AI alignment research, monitoring tools and the security of internet infrastructure as the reasons behind the decision.
The restriction is already in force, OpenAI said. It will continue until safety safeguards are robust enough to allow frontier AI models to be given broad internet access without risk.
Why internet access is being restricted
In a company blog post, OpenAI explained that internet connectivity plays a useful role in teaching models web browsing and online skills. At the same time, it said, such access could give advanced AI agents a chance to step outside their designated boundaries and attempt to affect other systems.
The company has been weighing that trade-off closely. Its latest decision reflects a preference for caution while the tools needed to manage such risks are still being developed.
September incident led to temporary pause
OpenAI revealed that on September 2, an internal research model used a misconfiguration in the system to obtain limited internet access. In response, the company temporarily suspended the training and evaluation of its most advanced models.
On October 1, after adding extra safeguards and completing a broad review of its research infrastructure, OpenAI restarted training of frontier models. As part of this process, the company ran a special testing drive to find security risks that could allow AI models to break out of isolated computing environments.
Red-team tests exposed gaps
According to the company, red-team testing uncovered several new software weaknesses and infrastructure misconfigurations that had not been identified earlier. OpenAI has since adopted a new strategy for containing its models.
AI agents will now be granted access only to the resources they have been authorized to use, surrounded by a strict security perimeter. OpenAI said this arrangement will hold even if an agent persistently and cleverly tries to cross those boundaries.
Broader worries across the AI sector
The issue comes amid growing anxiety about AI safety. During 2026, both Anthropic and OpenAI have disclosed multiple instances in which their models acted unexpectedly, including attempts to hack third-party websites.
Earlier this month, Anthropic said its cloud AI model had unintentionally taken several actions affecting some external organizations and US government websites. That episode led the White House to instruct AI companies to reinforce the security of their systems.