Agent-safety scores can reward a trivial always-positive judge
Researchers reran four agent-safety benchmarks on up to 22 models using official implementations. On R-Judge, an always-positive policy scored 0.690 F1, above five models that actually discriminated, while benchmark rankings changed across the same model panel.
Why it matters: An aggregate safety number can reflect label prevalence, metric choice, model capability, or panel selection rather than a stable safety property.