Frequently Asked Questions about MCP Server for Garak

Q: What is MCP Server? A: MCP (Model Context Protocol) Server is a lightweight server that facilitates communication between applications and Large Language Models (LLMs). It standardizes how applications provide context to LLMs, enhancing their functionality and security.

Q: What is Garak, and why is it important? A: Garak is an LLM vulnerability scanner. It helps identify potential weaknesses in AI models, preventing exploits that could lead to data breaches, model manipulation, or other security incidents. The MCP Server for Garak allows you to integrate Garak’s scanning capabilities into your workflow.

Q: What are the key features of the MCP Server for Garak? A: The key features include listing available attacks, running attacks on specified models, seamless integration with Garak, comprehensive attack library, automated scanning, detailed reporting, and a lightweight design.

Q: What are the prerequisites for installing the MCP Server for Garak? A: The prerequisites include Python 3.11 or higher and the ‘uv’ package installer. Optionally, if you plan to test Ollama models, the Ollama server should be running.

Q: How do I install the MCP Server for Garak? A: You can install it by cloning the repository from GitHub, navigating to the ‘src’ directory, and using ‘uv’ to manage dependencies and run the server.

Q: How do I configure the MCP Server for Garak with Cursor? A: Configuration for Cursor involves adding a ‘garak-mcp’ entry within the ‘mcpServers’ section of your Cursor settings, specifying the command to run the server along with its arguments and environment variables.

Q: What tools are provided by the MCP Server for Garak? A: The server provides tools such as list_model_types (lists available model types), list_models (lists models for a given type), list_garak_probes (lists available Garak attacks/probes), get_report (retrieves the last run’s report), and run_attack (executes an attack on a model).

Q: What model types are supported by the MCP Server for Garak? A: The server supports model types such as ollama, openai, huggingface, and ggml.

Q: What is UBOS, and how does the MCP Server for Garak integrate with it? A: UBOS is a full-stack AI Agent Development Platform focused on enabling businesses to create and deploy AI Agents. The MCP Server for Garak is available on the UBOS Asset Marketplace and integrates seamlessly with the UBOS platform, providing enhanced security for AI Agents.

Q: Where can I find the GitHub repository for the MCP Server for Garak? A: The repository is located at https://github.com/BIGdeadLock/Garak-MCP.git

Q: What future improvements are planned for the MCP Server for Garak? A: Future steps include adding support for Smithery AI (Docker and config), improving reporting capabilities, and testing/validating models from OpenAI, HuggingFace, and local GGML models.