ResearchBenchmarksLLM Safety Evaluation on Harmful Prompt DatasetFollow37.5Refusal Rate (Full)Claude22.00426.02730.0534.073Feb 10, 2026Evaluation ResultsMethodMethodLinksRefusal Rate (Full)ASR (Binary)WASRLeak Rate (Partial)Peak VulnerabilityClaude2026.0237.51542.819—GPT-52026.0226.225.855.98—Gemini2026.0222.626.959.54—