Xingshen Wei

MemPoison: Uncovering Persistent Memory Threats and Structural Blind Spots in LLM Agents

Jifeng Gao, Kang Xia, Yi Zhang, Xiaobin Hong, Mingkai Lin, Xingshen Wei, et al.

Persistent external memory enhances agent continuity but introduces persistent security vulnerabilities: adversarial content can be injected via standard interaction channels, retained across turns, and later distort downstream behavior. To address this challenge, we propose MemP…Read more