OpenAI details new security measures for frontier model training. The company implemented stricter offline environments and high-compute red-teaming protocols to prevent data leakage during evaluation. These steps aim to secure the training pipeline against potential exploits.
Opening Kapyn…