By Hamza Faisal
OpenAI has restarted training its most powerful AI models after an 11-day safety pause, and it is doing so under a strict new rule: no live internet access, ever, during frontier training. The company laid out the change in a blog post published October 23 titled "Securing our frontier model training environment," confirming that reinforcement learning training and internal inference of its frontier models resumed on October 1.
The pause began on September 20, when one of OpenAI's research models used a DNS misconfiguration to escape its offline environment and gain limited internet access. The company contained the incident, stopped frontier training, and used the gap to rebuild its containment. It then turned its most cyber-capable AI models loose on its own research infrastructure to see whether they could break out again.
The no-internet rule and the offline world
"Based on the current state of alignment research, monitoring systems, and the security of internet infrastructure, we believe the responsible approach is to conduct frontier model training with no internet access," OpenAI wrote. The rule is conditional, not permanent — it stays in place "until advances in these safeguards justify broader internet access."
Models still get to practice browsing, searching and using apps, but only inside a sealed setup. OpenAI built what it calls an "offline world": an indexed and cached copy of the web that agents can interact with without touching any live website. It is not a static text dump — models run browsers, computer-use tools and code environments inside the secure boundary, with communication restricted to authorized agents.
What the red-team agents found
The agents OpenAI set loose as attackers found real weaknesses. They discovered previously unknown flaws that let them escape the virtual machines meant to isolate them from the host system — vulnerabilities OpenAI says it has disclosed and fixed. In one test, agents placed on separate virtual machines with no network link still managed to communicate, passing messages through the processor's cache by exploiting the speed gap between cache and main memory. One agent pushed data out of the cache while another timed how long reloading took, recovering the message from the delay.
To address this, OpenAI added monitoring checks specifically designed to spot when models try to collaborate. The company says it has reserved substantial computing power for continuous red-teaming, and that high-compute red-teaming is now part of the safety cases that decide whether training and evaluation runs can go ahead at all.
A Brotli flaw and the outside verdict
The exercise produced a find far beyond OpenAI's own walls: a flaw in Brotli, the widely used data-compression tool. OpenAI said attackers could exploit it by sending malicious data that lets them run their own code on affected systems, including popular web software such as Node.js, Bun and Envoy. The root cause had quietly been fixed in Brotli's source code back in February, but no new release shipped at the time; updated versions and a public record followed OpenAI's disclosure.
Separately, AI safety evaluators METR and Redwood Research said they found several new vulnerabilities in OpenAI's internal systems in a pilot review earlier this year, including an exploit agents could have used to keep hundreds of their actions hidden from monitoring. OpenAI told the organizations it has since patched all vulnerabilities affecting its production safeguards, though METR and Redwood said they had not yet verified the fixes.
By Hamza Faisal



