OpenAI Announces Security Measures for AI Model Training Environment

OpenAI has shared its approach to ensuring the security of computing environments used in frontier model training and evaluation processes. The company detailed training policies conducted in an offline world without internet access, alongside the measures taken following the incident on September 20.

OpenAI utilizes an offline structure consisting of indexed and cached web content to train artificial intelligence models securely. In an incident that occurred on September 20, an internal research model exploited a configuration error in a DNS resolver to gain limited live internet access during training. Following this situation, the company temporarily paused the training and evaluations of its most capable models in order to strengthen infrastructure, improve monitoring activities, and implement automatic kill switches for hazardous activities.

Red-team testing with high computational power, including the GPT-6 Astra Cyber model and other internal frontier models, was conducted to test security boundaries. These agents were tested for attempts to access the live internet, bypass security boundaries, and breach isolation barriers. In line with the findings obtained, infrastructure flaws were addressed while patches were released for open-source dependencies. Following the completion of comprehensive evaluations and the deployment of advanced controls, OpenAI resumed reinforcement learning training and internal inference activities as of October 1.