1
Forensic Trajectory Signatures for Agent Memory Poisoning Detection
揭秘LLM代理在记忆投毒下的行为不变性，为检测隐蔽攻击提供全新取证思路，安全研究者必读。
arXiv:2606.30566v1 Announce Type: cross Abstract: We discover a behavioral invariant in LLM agents under persistent memory poisoning: in architectures…