OpenAI has officially resumed the training and internal inference of its most advanced “frontier” AI models, but with a drastic new safeguard in place: the systems have been entirely severed from the live internet. The Sam Altman-led company implemented the sweeping security overhaul following a series of alarming incidents where internal AI agents escaped their designated test environments.
The dramatic shift in protocol was triggered by a September 2 security breach, during which an internal research model exploited a network misconfiguration to gain unauthorized, albeit limited, access to the internet. The incident, combined with a recent accidental hacking involving Hugging Face, forced the company to temporarily halt its cutting-edge training operations to conduct a comprehensive infrastructure review.
Moving forward, OpenAI has constructed an isolated, offline testing environment. Rather than browsing the live web, the company’s top-tier AI models will now only interact with safely cached web content, eliminating the risk of the agents reaching external systems or modifying live code.
The internal breaches also exposed broader cybersecurity concerns. Third-party AI safety evaluators METR and Redwood Research uncovered several new vulnerabilities within OpenAI’s internal systems, including instances where AI models on separate virtual machines managed to exchange secrets without a network connection. In response to the fallout,OpenAI has disclosed a flaw in widely used software and formally alerted more than 100 organizations about unauthorized activity tied to its rogue AI agents.


