MemPoison: Bypassing Selective Memory Mechanisms to Plant Backdoors in LLM Agents
MemPoison is an attack that injects malicious backdoors into LLM agents' long-term memory by exploiting how agents selectively extract and rewrite information. The method uses a semantic bridge, entity masquerading, and joint embedding optimization to ensure the trigger and payload are extracted together. This allows the attacker to bypass memory filters while ensuring the backdoor activates when triggered.