ResearchBenchmarksJailbreak Attack Robustness on HarmBench 200 standard behaviorsFollow34.5GCG Success RateInitial1.229.8618.527.14Jun 2, 2026Evaluation ResultsMethodMethodLinksGCG Success RatePAIR Success RateTAP Success RateAutoDAN Success RateInitialModel=Llama-2-7b-Chat,...Model=Llama-2-7b-Chat, Alignment=Standard/Raw2026.0634.57.55.550InsertionModel=Llama-2-7b-Chat,...Model=Llama-2-7b-Chat, Alignment=Adversarial safety alignment2026.062.5110