Garak: the open source vulnerability scanner for LLMs

In traditional security there are scanners like Nessus or Nmap that automatically test a system for known vulnerabilities. Nothing equivalent existed for language models until Garak.

Garak works like an automated red teamer: it takes an LLM as a target, sends it hundreds of prompts designed to surface problematic behaviors, and produces a report with results classified by risk type.

It tests categories like the tendency to produce harmful content, resistance to jailbreak attempts, propensity to hallucinate on verifiable facts, and robustness against common prompt injection techniques. It is open source and works with OpenAI, Hugging Face, and local models via Ollama.