Reproduce “MemHarmBench: Evaluating Cross-Session Prompt Injection in Memory-Persistent AI Agents” on OpenResearch
Turn your favorite coding agents into research agents that review literature, develop hypotheses, and run experiments.
Universal app for Apple silicon and Intel. Requires macOS 11 or later.
About this paper
Zhihua Wang, Yunhao Feng, Yifeng Gao, Rui Pan, Baiheng Zhao, Zixing Chen, Xingjun Ma
The authors introduce MemHarmBench to test whether instructions encountered in external content persist in memory and change decisions. The benchmark removes the attack content before a new-session task, then measures target adoption, task completion, and execution anomalies. Across four configurations, 55.71% of eligible triggers combined adoption with task completion and no judged anomaly; this is not an all-case or real-world harm rate.
Summary by alphaXiv.
Tip: change arxiv.org to autoarxiv.org in any arXiv abstract link to open the paper here.