May 16, 2026 Red Teaming Language Models to Reduce Harms: Methods, Scaling Behaviors, and Lessons Learned