arXiv cs.AI· Jinhao Hu, Bardia Mohammadi, Mofasshara Binte Rafique, Lars Henning Klein, Akhil Arora, Ashvin Goel, Laurent Bindschaedler·· 12 小时前AI 评分38
ClawMMU：为有状态 LLM 智能体打造保护一致的虚拟内存
ClawMMU: Protection-Coherent Virtual Memory for Stateful LLM Agents
AI 导读
ClawMMU 是一个为有状态 LLM 智能体设计的内存保护层，为每段记忆标注所有者、作用域、来源与权限，并在内容进入模型上下文、写入存储、从存储返回或驱动工具调用时进行检查。
正文 · 原文
Abstract:AI agents keep long-term memory, much of which comes from web pages and tool results. Agent harnesses, the software that runs the model, do not record who may act on this memory, so an instruction hidden in a fetched page can be treated as the user's own instruction. We propose ClawMMU, a protection layer that labels every piece of memory with its owner, scope, origins, and authority. ClawMMU checks these labels whenever content enters the model's context, is written to storage, returns from storage, or drives a tool call. Its central rule is that information may rise in authority, but authority never lifts itself. Because the checks depend on where content came from, not on what it says, memory attacks become classical protection faults that the harness stops even when the model is fooled.
| Comments: | Accepted at the AgenticOS Workshop @ NeurIPS 2026 (non-archival). 6 pages, 1 figure, 1 table |
| Subjects: | Cryptography and Security (cs.CR); Artificial Intelligence (cs.AI); Operating Systems (cs.OS) |
| Cite as: | arXiv:2610.28093 [cs.CR] |
| (or arXiv:2610.28093v1 [cs.CR] for this version) | |
| https://doi.org/10.48550/arXiv.2610.28093 arXiv-issued DOI via DataCite (pending registration) |
Submission history
From: Jinhao Hu [view email]
[v1]
Tue, 20 Oct 2026 12:09:53 UTC (14 KB)
来源：arXiv cs.AI · arxiv.org