Reproduce “DrunkAgent: Stealthy Memory Corruption in LLM-Powered Recommender 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
Shiyi Yang, Zhibo Hu, Xinshu Li, Chen Wang, Tong Yu, Xiwei Xu, Liming Zhu, Lina Yao
DrunkAgent, a black-box attack framework, exploits memory vulnerabilities in LLM-powered recommender agents to stealthily promote specific items. This framework demonstrates high transferability across diverse agentic systems and datasets, outperforming existing baselines while maintaining linguistic naturalness.
Key insights
- LLM-powered agentic recommender systems are susceptible to memory-based adversarial attacks, revealing a fundamental vulnerability in their adaptive learning mechanisms.
- Attacks crafted with semantic stealth and dynamic perturbation strategies can persistently corrupt agent memories, proving more effective than general text perturbations.
- While advanced architectures like retrieval-augmented systems offer some resilience, they are not immune to such memory-centric black-box attacks.
Summary by alphaXiv.
Tip: change arxiv.org to autoarxiv.org in any arXiv abstract link to open the paper here.