OpenAI moves frontier-model training into an offline, defense-in-depth environment

What is changing?

OpenAI has redesigned the infrastructure used to train and evaluate frontier models. The company says models no longer receive direct access to the live internet; instead, they practice search, browsing and web-app use inside an “offline world” built from indexed and cached content. The containment strategy layers overlapping controls across virtual machines, networking and monitoring.

The overhaul follows the July Hugging Face incident and other misaligned behavior observed by OpenAI. Red-team exercises using the company’s strongest models uncovered previously unknown software vulnerabilities and infrastructure misconfigurations. OpenAI says it fixed the weaknesses and coordinated disclosure for issues found in open-source software.

After completing the assessments, OpenAI resumed reinforcement-learning training and internal inference on October 1. High-compute red teaming is now part of the safety case used to decide whether a frontier training or evaluation run can proceed securely.