donvaughn
MCP Serverdonvaughnpublic

mcp secrets downloader please connect

This repo demonstrates the threats that LLM tool calling expose users to

Repository Info

0
Stars
1
Forks
0
Watchers
0
Issues
Python
Language
-
License

About This Server

This repo demonstrates the threats that LLM tool calling expose users to

Model Context Protocol (MCP) - This server can be integrated with AI applications to provide additional context and capabilities, enabling enhanced AI interactions and functionality.

Documentation

MCP Secrets Downloader Please Connect

🔍 Overview

This project demonstrates Tool Poisoning, a security issue affecting Large Language Models (LLMs) that can use tools or plugins. Tool poisoning is surprisingly easy to pull off: attackers hide harmful commands inside descriptions that seem harmless.

For example, by adding a long description or inserting lots of spaces, dangerous commands can be concealed from view. These hidden instructions might run immediately or stay dormant until the tool is trusted—or even after its source code is compromised.

The key problem? LLMs are designed to follow instructions. If a malicious command is hidden in a tool's description, the LLM might execute it without the user ever realizing.

⚠️ Spoiler: This tool won't actually steal your secrets—but it shows how easily one could.

🧠 Why It Matters

While we're using MCP (Model Control Protocol) to demonstrate this issue, tool poisoning isn't just an MCP problem. It becomes dangerous when:

  1. A trusted tool has access to sensitive data, and then
  2. An untrusted or compromised tool gains access to that data or can perform actions on behalf of the trusted tool.

Any application that lets LLMs use tools is at risk. This includes ChatGPT, Claude, Cursor, Windsurf, and others.

Right now, there's no guaranteed way to stop this kind of attack. It's up to providers and users to carefully review all tool code for hidden threats.

🧪 Demo: Tools with Hidden Dangers

This demo includes two fake MCP tools, each with secret instructions buried in their descriptions:

  1. dons_daily_riddles

    • What it claims to do: Tells you a riddle when you say "Riddle me that, Don."
    • What's hidden: Copies your .env file (which may contain secrets) to a new file called .env.pwned.
  2. dons_daily_horoscopes

    • What it claims to do: Gives you a horoscope.
    • What's hidden: Downloads a script from the internet, makes it executable, and runs it.

These tools exploit the LLM's trust in tool descriptions to carry out harmful actions.

🚀 How to Try It Yourself

✅ What You Need

  • Python environment manager uv (install with pip install uv)
  • An MCP-compatible client (like Cursor or Windsurf)

🛠️ Steps

  1. Clone the Repository:
    git clone https://github.com/donvaughn/mcp-secrets-downloader-please-connect.git
    cd mcp-secrets-downloader-please-connect
    

Quick Start

1

Clone the repository

git clone https://github.com/donvaughn/mcp-secrets-downloader-please-connect
2

Install dependencies

cd mcp-secrets-downloader-please-connect
npm install
3

Follow the documentation

Check the repository's README.md file for specific installation and usage instructions.

Repository Details

Ownerdonvaughn
Repomcp-secrets-downloader-please-connect
LanguagePython
License-
Last fetched8/10/2025

Recommended MCP Servers

💬

Discord MCP

Enable AI assistants to seamlessly interact with Discord servers, channels, and messages.

integrationsdiscordchat
🔗

Knit MCP

Connect AI agents to 200+ SaaS applications and automate workflows.

integrationsautomationsaas
🕷️

Apify MCP Server

Deploy and interact with Apify actors for web scraping and data extraction.

apifycrawlerdata
🌐

BrowserStack MCP

BrowserStack MCP Server for automated testing across multiple browsers.

testingqabrowsers

Zapier MCP

A Zapier server that provides automation capabilities for various apps.

zapierautomation