garak

garak is a vulnerability scanner specifically designed for Large Language Models, developed by NVIDIA. It probes LLMs for weaknesses including prompt injection, jailbreaking, training data leakage, hallucination, toxic generation, and other failure modes. garak ships with dozens of probe modules targeting specific vulnerability classes and supports custom probe development. It works with OpenAI, Hugging Face, local models, and any API-compatible endpoint. Results include detailed reports on which attacks succeeded, confidence scores, and categorization by risk type. Essential for red-teaming AI systems before deployment and validating safety guardrails.