OpenAI plans to restrict internet access for its frontier models during training until adequate alignment, monitoring, and security safeguards are established (Rachel Metz/Bloomberg)

OpenAI has implemented a rigorous new security protocol mandating all frontier AI models to undergo training and evaluation without internet access. This policy change arises from the escalating risk that powerful AI systems could exploit software vulnerabilities, bypass digital barriers, and potentially interact with external systems in unforeseen ways.

The directive stems from an incident in late September where an internal research model inadvertently gained brief internet access during a training session, thanks to a misconfigured DNS resolver. Though the breach was swiftly contained, it prompted OpenAI to pause all advanced training and evaluation to overhaul its security infrastructure.

To enable models to retain crucial skills like web browsing and navigating intricate applications, researchers created an extensive offline environment. This secure, isolated digital realm relies solely on cached and indexed web data, eliminating the risk of external exposure while still allowing models to learn. To enforce this stringent isolation, OpenAI has adopted a multi-layered defense mechanism.

At the software level, highly restricted containers reside within lightweight virtual machines. These localized defenses coexist with cloud-level firewalls, identity management protocols, and automated monitoring systems, all designed to instantly halt any process displaying high-risk or unauthorized behavior. Before resuming advanced reinforcement learning in October, OpenAI subjected its new infrastructure to a simulated attack from its own AI systems.

Advanced models, including the forthcoming GPT-6 Astra Cyber, were tasked as aggressive red-teaming fleets, given access to system source code to deliberately search for weaknesses. The AI fleets successfully identified previously unknown vulnerabilities and misconfigurations, which were promptly addressed. Going forward, OpenAI will mandate continuous, high-compute red-teaming to ensure its containment architecture remains robust against the capabilities of future models.