Reproduce “Cowpox: Towards the Immunity of VLM-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
Yutong Wu, Jie Zhang, Yiming Li, Chao Zhang, Qing Guo, Nils Lukas, Tianwei Zhang
Researchers from NTU, A*STAR, Tsinghua, and MBZUAI developed COWPOX, a distributed defense framework that establishes immunity against infectious jailbreak attacks in Vision-Language Model (VLM)-based multi-agent systems. It achieves over 95% recovery of infected agents by autonomously generating and propagating 'cure samples' that outcompete malicious content via RAG mechanisms, even with minimal defender control.
Key insights
- Framing AI system security against spreading attacks as an 'epidemic' and developing an 'immunity' mechanism introduces a novel epidemiological perspective to AI security.
- Manipulating Retrieval-Augmented Generation (RAG) scores is an effective method to make benign 'cure samples' outcompete malicious virus samples, thereby halting infection spread and promoting recovery.
- A theoretical framework, including Proposition 4.1, establishes that full system recovery is guaranteed if cured agents convert infected agents faster than infected agents convert cured ones.
Summary by alphaXiv.
Tip: change arxiv.org to autoarxiv.org in any arXiv abstract link to open the paper here.