The harmful and benign prompts were sourced from a Cornell University dataset designed to rigorously test AI security, drawing from established red-teaming methodologies. While not a reasoning-based ...