Reproduce “Collaborative Shadows: Distributed Backdoor Attacks in LLM-Based Multi-Agent Systems” 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
Pengyu Zhu, Lijun Li, Yaxing Lyu, Li Sun, Sen Su, Jing Shao
Researchers at Shanghai Artificial Intelligence Laboratory and Beijing University of Posts and Telecommunications introduce "Collaborative Shadows," the first distributed backdoor attack for LLM-based multi-agent systems, demonstrating how inter-agent collaboration can be exploited. This attack achieves a consistent Attack Success Rate over 95% across various LLMs and benchmarks, with minimal impact on benign task completion.
Key insights
- Inter-agent collaboration itself constitutes a novel and potent attack surface, allowing for the distributed assembly and execution of backdoors across multiple agents.
- Distributed backdoor attacks can be highly effective and stealthy, achieving near-perfect activation rates without disrupting the benign functionality of the multi-agent system.
- The attack's success relies on the synergistic combination of both poisoned tools and precisely engineered attack-variant instructions, with accidental triggering being highly improbable under normal operation.
Summary by alphaXiv.
Tip: change arxiv.org to autoarxiv.org in any arXiv abstract link to open the paper here.