Safety Benchmark Reveals Discrepancies Between Frontier and Specialized Models

An evaluation of 53 models across 11 datasets shows that frontier models often fail in specific harm categories where smaller, specialized moderators excel. The study highlights critical blind spots in conversational safety and the inadequacy of a single-model safety layer.