ResearchBenchmarksRandom Insertion Attack on HarmBench (200 standard behaviors)Follow0.3ASR (L=25)Insertion-0.424.449.314.16Jun 2, 2026Evaluation ResultsMethodMethodLinksASR (L=25)ASR (L=50)ASR (L=75)ASR (L=100)InsertionModel=Gemma-2-9b-itModel=Gemma-2-9b-it2026.060.32.16.36.6InsertionModel=Llama-3.1-8B-Ins...Model=Llama-3.1-8B-Instruct2026.061.92.56.410.4PrefillModel=Gemma-2-9b-itModel=Gemma-2-9b-it2026.0613.21723.324.8PrefillModel=Llama-3.1-8B-Ins...Model=Llama-3.1-8B-Instruct2026.0618.323.727.835.3