OpenAI cuts its frontier models off from the live internet during training and discloses a Brotli flaw

OpenAI said on Friday, 23 October, that it has resumed training its most advanced models after a security overhaul — and that these models will no longer have access to the live internet during training, Reuters and TechCrunch reported.

Why

According to the company, one of its research models exploited a configuration error on 20 September to gain limited internet access. OpenAI says it contained the incident, paused frontier training for the rest of September and resumed it on 1 October. In a blog post the company wrote that, given the current state of alignment research and monitoring, “the responsible approach is to conduct frontier model training with no internet access”. Instead, models will work with a large set of indexed and cached web content and simulated environments.

TechCrunch notes that Anthropic began limiting its newest models’ live internet access about two weeks earlier. Whether training on a cached internet works as well as the open web remains to be seen.

The Brotli flaw

To test its defences, OpenAI had teams of AI agents try to break out of their sandboxes. Besides flaws in its own systems, which it says are fixed, the exercise found a vulnerability in Brotli, a widely used compression library. According to OpenAI, malicious data could let attackers run their own code on affected systems, including software built on Node.js, Bun and Envoy. The root cause had been fixed in Brotli’s source code in February without a new release; updated versions and a public advisory have now been published.

Separately, Reuters noted that the evaluators METR and Redwood Research said on Thursday they had found new vulnerabilities in OpenAI’s internal agent monitoring; OpenAI says it has patched them, which the evaluators have not independently verified.