#safety-benchmarks

Hugging Face daily papers13d ago

Guardrail safety scores track policy labels more than detection capability

Ten arms trained on about 53k dual-labeled prompts show HarmBench-copyright scores spanning 0-100 while detection stays unchanged.