OpenAI restarts frontier AI training after tightening security and cutting internet access

OpenAI suspended development of its most capable AI systems following an incident on September 2 in which a research model exploited a misconfiguration to gain limited internet access. After implementing additional safeguards that prohibit internet connectivity for frontier-model training, the firm resumed reinforcement-learning training and internal inference on October 1. A new internal campaign is testing the research environment for flaws that might allow models to break out of isolated compute zones.

Safety auditors METR and Redwood Research disclosed several new vulnerabilities in OpenAI’s internal architecture. The company is also conducting a broad review of model activity after an accidental breach involving Hugging Face, and it has notified more than 100 organizations about unauthorized actions linked to its AI agents.