MemPoison: Bypassing Selective Memory Mechanisms to Plant Backdoors in LLM Agents
Large language model (LLM) agents increasingly leverage long-term memory to support persistent and autonomous task execution. However, this capability also introduces a new attack surface: memory poisoning, where adversaries can inject malicious information to influence future behavior. Existing memory poisoning attack
研究重點
Large language model (LLM) agents increasingly leverage long-term memory to support persistent and autonomous task execution. However, this capability also introduces a new attack surface: memory poisoning, where adversaries can inject malicious information to influence future behavior. Existing memory poisoning attack
為什麼值得關注
這個議題涉及 Crew Pocket、Crew Deck 的實作方向。可先評估現有做法,再決定是否安排小型實驗。
可以進一步驗證什麼?
- 閱讀原文,確認它解決的問題、實驗方法及適用限制
- 檢查相關功能現有做法,找出可比較的差異
- 用小型 PoC 驗證成功率、品質、延遲和實際成本
參考資料
MemPoison: Bypassing Selective Memory Mechanisms to Plant Backdoors in LLM Agents ↗2026-11-12T00:00:00Z · 研究論文索引
延伸探索
相關研究
- VANTAGE-V-AI: Bounded-Memory KV-Cache Compression and Streaming Inference Runtime for Large Language Models相同研究主題 · 2026-11-17
- Jacobian-Causal Memory Control for Persistent Memory in Language Model Systems相同研究主題 · 2026-11-13
- Durable Memory on a Fixed Context Budget: Architecture and Measurements of a Long-Lived Coding Agent相同研究主題 · 2026-11-13