LLM Agent Memory Vulnerable to Manipulation, Study Finds

AI agents, which extend large language models with capabilities like tool use and memory, introduce new security vulnerabilities. This research specifically investigates how an agent's memory component can be exploited. By injecting corrupted or misleading information into an agent's memory, its subsequent decision-making can be compromised. The study focused on multiple-choice question answering, showing that even simple memory alterations led to the agent selecting incorrect answers despite receiving valid questions. This highlights a critical attack surface in the design and deployment of LLM-based agents, emphasizing the need for robust security measures.