2026. Evaluating generative AI safety on real-world harm: A focused classification study using the RealHarm dataset. AI and Wisdom. 1, 2 (Jun. 2026).