Paper proposes signed private retrieval against agent memory poisoning
Original title: On Private Knowledge Retrieval and Private Inference for secure multi-agent systems
The design uses a local Horn-clause reasoner and an externally signed knowledge base with k-anonymity or IT-PIR.
The paper examines semantic-memory poisoning in LLM-based multi-agent systems, where a malicious agent changes stored knowledge and therefore behavior. It proposes private knowledge retrieval and private inference, plus signatures on an external knowledge base, so agents can reject undesired manipulation. A local Prolog-like Horn-clause inference system queries that signed base privately. Python implementations cover k-anonymity and information-theoretical private information retrieval.
Unconfirmed: The paper discusses memory poisoning and a proposed design; it does not report a confirmed incident.
- The focus is semantic memory, not episodic or short-term memory.
- Knowledge is represented with Horn clauses.
- Privacy options include k-anonymity and information-theoretical PIR.
- Implementations are described in Python.
From the original article
1 min read at arxiv.orgMulti-agent systems provide a convenient framework for the massive deployment of agents based on Large Language Models (LLMs). It provides a sound framework where agents can interact with existing software to perform actions in an autonomous way. Agentic AI is the new term to denote these AI-based autonomous systems.
By Vicenç Torra · arxiv.org
A short excerpt; the article and its images belong to the publisher. Summary, key facts and assessment above are AI-written from the article and can be wrong.