MCP Server For Garak LLM Vulnerability Scanner

A lightweight MCP (Model Context Protocol) server for Garak.

Example:

https://github.com/user-attachments/assets/f6095d26-2b79-4ef7-a889-fd6be27bbbda

---

Features

• List Attacks: List all the attack available on Garak.
• Run Attack: Run the attack on a given model.

---

Prerequisites

1. Python 3.11 or higher: This project requires Python 3.11 or newer.

   # Check your Python version
   python --version
   

2. Install uv: A fast Python package installer and resolver.

   pip install uv
   

   Or use Homebrew:

   brew install uv
   

3. Optional: Ollama: If you want to run attacks on ollama models be sure that the ollama server is running.

ollama serve

---

Installation

Clone this repository:

git clone https://github.com/BIGdeadLock/Garak-MCP.git
cd src

---

Configuration

For Cursor users:

{
  "mcpServers": {
    "garak-mcp": {
      "command": "uv",
      "args": ["--directory", "path-to/Garak-MCP", "run", "garak-server"],
      "env": {}
    }
  }
}

---

Tools Provided

Overview

Name	Description
list_model_types	List all available model types (ollama, openai, huggingface, ggml)
list_models	List all available models for a given model type
list_garak_probes	List all available Garak attacks/probes
get_report	Get the report of the last run
run_attack	Run an attack with a given model and probe

Detailed Description

• list_model_types

  • List all available model types that can be used for attacks
  • Returns a list of supported model types (ollama, openai, huggingface, ggml)

• list_models

  • List all available models for a given model type
  • Input parameters:
    • model_type (string, required): The type of model to list (ollama, openai, huggingface, ggml)
  • Returns a list of available models for the specified type

• list_garak_probes

  • List all available Garak attacks/probes
  • Returns a list of available probes/attacks that can be run

• get_report

  • Get the report of the last run
  • Returns the path to the report file

• run_attack

  • Run an attack with the given model and probe
  • Input parameters:
    • model_type (string, required): The type of model to use
    • model_name (string, required): The name of the model to use
    • probe_name (string, required): The name of the attack/probe to use
  • Returns a list of vulnerabilities found

---

Future Steps

• [ ] Add support for Smithery AI: Docker and config
• [ ] Improve Reporting
• [ ] Test and validate OpenAI models (GPT-3.5, GPT-4)
• [ ] Test and validate HuggingFace models
• [ ] Test and validate local GGML models