From Untrusted Input to Trusted Memory: A Systematic Study of Memory Poisoning Attacks in LLM Agents

This study presents a systematic analysis of memory poisoning attacks in LLM-based agents, identifying four memory write channels and nine vulnerabilities related to model capabilities and architecture. A new benchmark, MPBench, is introduced to evaluate these attacks, revealing that agents with aggressive memory write and retrieval capabilities are more susceptible. The research highlights the inadequacy of current prompt injection defenses against memory poisoning, emphasizing the need for enhanced security measures in AI agent memory management.