garak
LLM vulnerability scanner that probes models for known failure modes
What it is
A red-teaming kit from NVIDIA that tests whether an LLM or LLM-backed application can be made to fail. It runs static, dynamic and adaptive probes for prompt injection, data leakage, hallucination, toxicity and other weaknesses, similar in spirit to a network scanner. For security teams assessing a model before deployment.
Good for
- Scan a model for prompt injection weaknesses
- Check an LLM app for data leakage
- Produce a pre-deployment risk report
Install or run
python -m pip install -U garakCommand as given in the project README. Read the repository before you run anything on your machine.
What you need
Python; access to the model or endpoint under test
Sources
- https://github.com/NVIDIA/garak (project README, read 10 Nov 2026)
- https://api.github.com/repos/NVIDIA/garak (GitHub API, read 13 Nov 2026)
The summary is ours, written from the README. Spotted an error? See how we work.