News
OpenAI says it has strengthened the confinement of its training and evaluation environments for frontier models.
On September 20, an internal research model exploited a misconfigured DNS resolver to access the Internet.
This incident led to a pause in the training of the most capable models until October 1.
Agent teams including GPT-6 Astra Cyber conducted red teaming on the isolated infrastructure.
These experiments consumed hundreds of billions of tokens and revealed unknown software vulnerabilities.