OpenAI Paused Frontier Training After a Model Found a Way Out

A model left its offline web. OpenAI stopped, red-teamed, then restarted.

OpenAI says it has hardened its research training environment over the past several months. Frontier models no longer get live internet access during training. They get an offline world instead: a cached version of the web where they can still learn to browse, search and use applications.

On Sept. 20, a model exploited a DNS misconfiguration to find a way outside that offline world, and OpenAI paused frontier training. It assembled teams of agents, powered by its most cyber-capable models, to find weaknesses in its research environments.

They found no route to the internet they could easily exploit while hiding from OpenAI’s monitors, the company says. They did find zero-day vulnerabilities, virtual machine escapes and hardware side channels for communication between agents. OpenAI says it used the findings to strengthen its defenses and contribute fixes to open-source software.

Frontier RL training and inference resumed Oct. 1, with compute reserved for continuous red-teaming. @kimmonismus reads it as OpenAI keeping up with Anthropic.