Garak
Framework for adversarial testing and evaluation of language models.
Made by NVIDIARed teaming and evaluation
About Garak
garak is an LLM vulnerability scanner. Use it to assess a model or system's security. garak is currently backed by NVIDIA and will continue to be.
garak is open-source software and actively updated, by both NVIDIA and the community. The developers are active in the garak discord.
Drafted from garak.ai · not yet confirmed by NVIDIA
How it works
garak is NVIDIA's vulnerability scanner for LLMs, modeled on nmap. Point it at a model and it runs probes (prompt injection, jailbreaks, data leakage, toxicity, hallucination, encoding attacks, known exploits) and detectors that check the responses, then reports what worked. Probes are organized by attack class and it's easy to add new ones.
It's a scanner rather than a framework: less flexible than PyRIT, faster to get a first report.
From the project
From the project
Reviews
Be the first to review it
AnonymousKnow a team weighing Garak? Send them this page.