Memory poisoning attacks on retrieval-augmented Large Language Model agents via deceptive semantic reasoning

Jing Hao

doi:10.1016/j.engappai.2026.113968

Memory poisoning attacks on retrieval-augmented Large Language Model agents via deceptive semantic reasoning

2026-01-29

0

PRE

AI

DOI：10.1016/j.engappai.2026.113968

Original

Original request for help

Save

Abstract

En 中文

• We propose a security evaluation framework to evaluate LLM-based agents under memory poisoning attacks. • We introduce a two-stage optimization using self-refine and chain-of-thought reasoning. • Experiments show high attack success and strong generalization across models and retrievers. • We uncover critical vulnerabilities in memory-augmented LLM agents against stealthy attacks.

Memory poisoning attacks on retrieval-augmented Large Language Model agents via deceptive semantic reasoning

Abstract

Journal

Organization