MCP Server For Garak LLM Vulnerability Scanner

@EdenYavin

Basic information

About MCP Server For Garak LLM Vulnerability Scanner

MCP Server for using Garak LLM vulnerability scanner

MCP Server For Garak LLM Vulnerability Scanner config

Add this server to your MCP-compatible client using the configuration below.

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

MCP Server For Garak LLM Vulnerability Scanner tools

Tool list not available yet

This listing does not currently have a tool list to display. This does not mean the server has no tools; check its documentation for supported capabilities.

We auto-extract tools from the README. The maintainer can list them under a ## Tools heading to populate this section.

Overview

What is MCP Server For Garak LLM Vulnerability Scanner?

A lightweight MCP server that integrates the Garak LLM vulnerability scanner, enabling AI assistants to list model types, models, probes, run attacks, and retrieve reports. It is intended for developers and security researchers assessing LLM robustness.

How to use MCP Server For Garak LLM Vulnerability Scanner?

Install prerequisites (Python 3.11+, uv), clone the repository, then configure your MCP host (e.g., Cursor, Claude Desktop) with the provided JSON entry pointing to the garak-server script via uv run. Invoke tools like run_attack with required parameters (model_type, model_name, probe_name).

Key features of MCP Server For Garak LLM Vulnerability Scanner

List available model types (ollama, openai, huggingface, ggml)
List all available models for a given type
List all available Garak probes/attacks
Run an attack with specified model and probe
Get the report path of the last run

Use cases of MCP Server For Garak LLM Vulnerability Scanner

Security audit of a local Ollama model against known probes
Automated vulnerability scanning of OpenAI or HuggingFace models
Batch testing of multiple probes on a single model
Integration with Cursor or Claude Desktop for on-demand scanning
Generating and retrieving detailed attack reports

Frequently asked questions

What prerequisites are needed?

Python 3.11 or higher, `uv` (install via `pip install uv` or Homebrew), and optionally Ollama running (`ollama serve`) for local models.

Where does the report data live?

The `get_report` tool returns the path to the report file from the last run; the README does not specify the exact directory.

Which MCP hosts are supported?

Tested on Cursor and Claude Desktop; configuration JSON is provided for these.

What transport or authentication is used?

The README does not mention transport or authentication details. It uses the MCP protocol, likely stdio transport via the `uv` command.

Are there known limitations?

Future steps include Docker/Smithery support, improved reporting, and validation for OpenAI, HuggingFace, and local GGML models (currently untested).

Comments

More AI & Agents MCP servers

AON — Agent Offer Network

AON

AON MCP connects AI agents to real, buyable products and services with live pricing and tracked links—enabling builders to earn developer revenue share from attributed clicks and conversions.

Quest: Hire a Human

Quest (hireahuman.quest)

Hire a verified local human for small real-world tasks your AI agent can't do: errands, queues, pickups, in-person checks and on-site photos. Your agent posts a quest, local humans send offers, and your user picks one an

q-ring

I4cTime

Quantum-inspired keyring for AI coding agents. Secure secrets with superposition, entanglement, tunneling, and teleportation — CLI + 44 MCP tools.

fhirHydrant

faulkj

Open-source Node.js FHIR MCP server with SMART Backend Services, metadata-aware search/CRUD tools, compact responses, FHIRPath filtering, safe pagination, audit events, and terminology lookup.

Factanker

FACTANKER serves 1.18B+ evidence-backed facts about US companies, banks, nonprofits and federal spending, sourced from SEC EDGAR, FFIEC call reports, IRS 990 filings and USAspending. Every value carries its evidence: fil