ShadowMerge: A Novel Poisoning Attack on Graph-Based Agent Memory via Relation-Channel Conflicts
ShadowMerge:一种通过关系通道冲突的图基agent内存新型污染攻击
机构 * Key Laboratory of Trustworthy Distributed Computing and Service (MoE)(可信分布式计算与服务重点实验室) ; Beijing University of Posts and Telecommunications(北京邮电大学) ; Zhongguancun Laboratory(中关村实验室)
专题命中 记忆与上下文管理 :agent(title,title_cn);分类 cs.AI
AI总结 本文提出ShadowMerge攻击,通过关系通道冲突攻击图基agent内存,利用冲突值影响agent行为,改进现有攻击方法,实现93.8%的攻击成功率。
Comments Preprint. Corresponding authors: Zifeng Kang and Tiantian Ji. Code is available at https://anonymous.4open.science/status/ShadowMerge-033C