Reproduce “Parasites in the Toolchain: A Large-Scale Analysis of Attacks on the MCP Ecosystem” 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
Shuli Zhao, Qinsheng Hou, Zihan Zhan, Yanhao Wang, Yuchong Xie, Yu Guo, Libo Chen, Shenghong Li and 1 more
A new class of privacy-leakage attacks, "Parasitic Toolchain Attacks" (MCP-UPD), is identified within the Model Context Protocol (MCP) ecosystem, where large language models are leveraged to orchestrate individually benign tools for stealthy data exfiltration. An analysis of public MCP platforms reveals 8.7% of tools and 27.2% of servers contain components exploitable in such attacks, highlighting systemic vulnerabilities.
Key insights
- Adversaries can leverage the autonomous orchestration capabilities of LLMs via MCP to chain individually benign tools (External Ingestion, Privacy Access, Network Access) for unauthorized privacy exfiltration.
- The core vulnerabilities enabling these attacks are the insufficient isolation between untrusted external content and LLM executable instructions, and the absence of least-privilege enforcement for tool permissions.
- Different large language models exhibit varying levels of susceptibility to these attacks, indicating that LLM alignment and safety measures play a critical role in resistance.
Summary by alphaXiv.
Tip: change arxiv.org to autoarxiv.org in any arXiv abstract link to open the paper here.