MemPot: Defending Against Memory Extraction Attacks with Optimized Honeypots

The researchers propose MemPot, which defends against memory extraction attacks by injecting optimized honeypots into the memory of large language model agents, and state that this is the first theoretically validated defense framework of this kind. The method uses two-stage optimization to generate trap documents that have a higher retrieval probability for attackers while being inconspicuous to normal users. The study models the detection process as a Wald sequential probability ratio test (SPRT) and proves that its average number of sampling rounds is lower than that of the optimal static detector. Experiments show a 50% improvement in detection AUROC, an 80% improvement in true positive rate under a low false positive rate constraint, no increase in online inference latency, and preservation of agent utility on standard tasks.