OpenAI explains how we strengthened containment for frontier model training and evaluation, including offline environments and high-compute red-teaming.