garak

Why we included this project

garak is a command-line scanner that probes LLM-backed systems to see how a model behaves under pressure. Point it at a Hugging Face checkpoint, a gguf file, or an OpenAI, Bedrock, or Replicate endpoint, and it tests for concrete failure modes like prompt injection, jailbreaks, data leakage, and toxicity, then reports a pass/fail rate for each probe. That output works both as a routine safety gate in CI before a model ships and as an on-demand red-teaming pass when you are evaluating a new vendor or open-weight model. The nmap-for-LLMs comparison is fair: it enumerates a target's exposures rather than monitoring or defending it, so it slots into existing audit workflows without requiring a heavy platform. Security engineers and app developers who want evidence of a model's weak spots before users find them will get practical use out of it.