Reproduce “CORBA: Contagious Recursive Blocking Attacks on Multi-Agent Systems Based on Large Language Models” 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
Zhenhong Zhou, Zherui Li, Jie Zhang, Yuanhe Zhang, Kun Wang, Yang Liu, Qing Guo
This research introduces Contagious Recursive Blocking Attacks (CORBA), a novel attack vector that significantly disrupts the availability and consumes computational resources of Large Language Model-based Multi-Agent Systems (LLM-MASs). CORBA leverages seemingly benign instructions to achieve high attack success rates and rapid propagation across various system configurations, while effectively evading existing LLM safety mechanisms.
Key insights
- LLM-MASs are highly vulnerable to contagious recursive blocking attacks, which can severely degrade system availability and exhaust computational resources.
- CORBA's combination of recursive persistence and contagious propagation allows it to spread efficiently and effectively across diverse multi-agent system frameworks and network topologies.
- Existing defense mechanisms, including LLM-based safety checkers, dynamic monitors, and perplexity-based detection, are largely ineffective at detecting or mitigating CORBA due to its use of benign-appearing instructions.
Summary by alphaXiv.
Tip: change arxiv.org to autoarxiv.org in any arXiv abstract link to open the paper here.