Memory Poisoning Attack and Defense on Memory Based LLM-Agents

Collection Summary

Query-only interaction으로 persistent memory를 오염시키는 MINJA-style attack의 현실적 조건을 평가하고 moderation과 trust-aware memory sanitization을 비교한다.

Rollout-Buffer Relevance

• **Target store**: pre-existing legitimate and injected long-term memories.
• **Defense candidates**: composite trust scoring, temporal decay, pattern filtering, trust-aware retrieval.
• **Security relevance**: shows that buffer occupancy, prior legitimate content, retrieval parameters, trust threshold calibration materially change attack effectiveness and utility loss.
• **Affected types**: memory-based assistants, healthcare agents, retrieval-driven experiential agents.