garak, LLM vulnerability scanner

Garak is an open-source LLM vulnerability scanner developed by NVIDIA that systematically probes large language models and dialogue systems for security weaknesses and failure modes. Garak brings a structured, adversarial testing methodology to the evaluation of language models.

The tool combines static, dynamic, and adaptive probes across a comprehensive range of attack vectors and vulnerability categories. These include prompt injection, jailbreaks, hallucination, toxicity generation, data leakage, misinformation, encoding-based attacks, malware generation, cross-site scripting, package hallucination, and glitch token exploitation, among many others. Each probe is paired with detectors that assess whether a model's output exhibits the targeted failure mode, and results are logged in structured JSONL reports that support downstream analysis and comparison across models and runs.

garak supports most if not all major LLM providers and interfaces, including OpenAI, Hugging Face, AWS Bedrock, Replicate, Cohere, Groq, NVIDIA NIM, gguf models, and any REST-accessible endpoint. This broad compatibility makes it applicable across research environments, enterprise deployments, and regulatory evaluation contexts without platform dependency.