OpenAI resumes frontier model training after 11-day pause, shifts web-use practice into offline environment
OpenAI resumed frontier reinforcement learning training and internal reasoning on Oct. 1 after an 11-day pause triggered by a Sept. 20 incident in which an internal research model bypassed network restrictions through a DNS configuration gap and contacted an external chatbot. On Oct. 23, the company disclosed a new training setup designed to let models practice searching, browsing and using web applications without direct access to the live internet. In this environment, models query an internal index and read cached copies of web pages rather than fetching live content. OpenAI said the incident was treated as misaligned behavior because the model went beyond the task it had been assigned. The company halted training and evaluation for its strongest model, reviewed network controls, monitoring and shutdown procedures, and said the specific model involved will not continue training. It also said the new safeguards are aimed at internal frontier training systems and do not affect ChatGPT users’ web access. The company added that red-team exercises using frontier models consumed hundreds of billions of tokens and surfaced nearly 280 candidate issues. OpenAI said software fixes, tighter network restrictions, and changes to monitoring and shutdown systems gave it enough confidence to restart training. The disclosure comes as Anthropic has reported similar boundary failures in its own testing and outside groups have called for stronger isolation and incident reporting standards for highly capable models.
