Tongyu Ge | Searcharxiv

arXiv2026

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

Memory is a core component of AI agents, enabling them to accumulate knowledge across interactions and improve performance. However, persistent memory introduces the risk of memory poisoning, where a single adversarial memory write can exert long-term influence over agent behavior. We present a systematic study of memory poisoning in LLM-based agents. We identify four memory write channels and nine structural vulnerabilities in model capabilities, system prompt design, and agent system architecture that make these channels exploitable. Based on these vulnerabilities, we develop a taxonomy of six classes of memory poisoning attacks. Furthermore, we design MPBench -- a benchmark for evaluating memory poisoning attacks, and show that agents designed to write and retrieve memory more aggressively are more exploitable. We also show that existing prompt injection defenses fail to cover memory poisoning attacks. Our findings provide a foundation for understanding and mitigating memory poisoning attacks against AI agents.

cs.CR↗

arXiv2025

Towards Privacy-Preserving Split Learning: Destabilizing Adversarial Inference and Reconstruction Attacks in the Cloud

This work aims to provide both privacy and utility within a split learning framework while considering both forward attribute inference and backward reconstruction attacks. To address this, a novel approach has been proposed, which makes use of class activation maps and autoencoders as a plug-in strategy aiming to increase the user's privacy and destabilize an adversary. The proposed approach is compared with a dimensionality-reduction-based plug-in strategy, which makes use of principal component analysis to transform the feature map onto a lower-dimensional feature space. Our work shows that our proposed autoencoder-based approach is preferred as it can provide protection at an earlier split position over the tested architectures in our setting, and, hence, better utility for resource-constrained devices in edge-cloud collaborative inference (EC) systems.

cs.CR↗