Traditional evaluations and red-teaming remain essential, especially for rare or severe risks. Deployment Simulation complements them by helping us estimate how often undesired behaviors may occur in realistic use and surface new behaviors before release.

→ View original post on X — @openai