2025-01-15Proj CJI Paper Reading: AdaPPA: Adaptive Position Pre-Fill Jailbreak Attack Approach Targeting LLMsAbstractBackground:目前的jailbreakmutator方式更集中在语义level，更容易被防御措施检查到本文:AdaPPA(AdaptivePositionPre-FilledJailbreakAttack)Task:adaptivepositionpre-filljailbreakattackapproachMethod:利用模型的instructionfollowing能力，先输出p