ResearchBenchmarksPrefill Attack on HarmBench Harmful 200 Standard Behaviors 13Follow13.8ASR (Length 25)Insertion0.84.1757.5510.925Jun 2, 2026Evaluation ResultsMethodMethodLinksASR (Length 25)ASR (Length 50)ASR (Length 75)ASR (Length 100)InsertionModel=Gemma-2-9b-itModel=Gemma-2-9b-it2026.0613.818.127.129.2PrefillModel=Gemma-2-9b-itModel=Gemma-2-9b-it2026.067.113.52222.6InsertionModel=Llama-3.1-8B-Ins...Model=Llama-3.1-8B-Instruct2026.061.64.210.923.1PrefillModel=Llama-3.1-8B-Ins...Model=Llama-3.1-8B-Instruct2026.061.34.220.629