Evaluating Generative AI Safety on Real-World Harm: A Focused Classification Study Using the RealHarm Dataset. AIW 2026, 1 (2).