MEMSAD: Gradient-Coupled Anomaly Detection Catches Memory Poisoning in RAG Agents

arXiv·medium signal

MEMSAD formalizes memory poisoning attacks on retrieval-augmented agents as a Stackelberg game and provides a unified evaluation framework spanning three attack classes. Persistent external memory lets agents maintain context across sessions but its security is formally uncharacterized until now. For builders running RAG agents with persistent memory stores, this is the first framework for detecting poisoned entries before they corrupt downstream reasoning — a growing attack surface as agents get long-term memory.