**garak** logo

**garak**

Free

![](https://img.shields.io/github/stars/NVIDIA/garak?style=flat-square)

FreeFree tier
Type
Open Source
Company
NVIDIA
LinksX

About **garak**

garak is a free, open-source command-line tool developed by NVIDIA for probing and assessing vulnerabilities in large language models (LLMs). It functions as a red-teaming assessment kit, similar to security tools like nmap or Metasploit Framework, but focused on LLMs. garak checks if an LLM can be made to fail in undesirable ways by probing for hallucination, data leakage, prompt injection, misinformation, toxicity generation, jailbreaks, and many other weaknesses. It combines static, dynamic, and adaptive probes and supports a wide range of LLM backends including Hugging Face, OpenAI, AWS Bedrock, Replicate, LiteLLM, and GGUF models. The tool is maintained by NVIDIA and has an active community with a Discord server and Twitter presence.

Key Features

Probes for hallucination, data leakage, prompt injection, misinformation, toxicity generation, and jailbreaks
Combines static, dynamic, and adaptive probes
Supports Hugging Face Hub models, OpenAI API, AWS Bedrock, Replicate, LiteLLM, GGUF models, and more
Command-line interface with options to list probes and specify target models and generators
Free and open-source (MIT license implied by GitHub repository)
Active development with community support on Discord and Twitter

Pros & Cons

Pros
  • Free and open-source with no usage restrictions
  • Comprehensive probe library covering many known LLM weaknesses
  • Supports a wide variety of LLM backends and interfaces
  • Active development and community contribution
  • Designed for both static and adaptive probing, similar to professional security tools
Cons
  • Command-line only, lacking a graphical user interface
  • Requires technical expertise to set up and configure (e.g., API keys, model names)
  • Documentation may be spread across multiple sources (GitHub, docs.garak.ai)
  • Some probes may require significant computational resources depending on the model size

Best For

Red-teaming and security auditing of large language modelsIdentifying vulnerabilities in LLM-based applications before deploymentCompliance and safety testing for generative AI systemsResearch into LLM failure modes and alignment weaknesses