> ## Content Index
> Fetch the complete content index at: https://www.humanoidthreats.com/llms.txt
> Use this file to discover other available public pages before exploring further.

# garak (NVIDIA)
- URL: https://www.humanoidthreats.com/directory/garak-nvidia/
- Published: 2026-11-01T07:26:05.000Z
- Updated: 2026-11-01T07:26:05.000Z
- Description: Open-source LLM vulnerability scanner maintained by NVIDIA and the community that probes models for jailbreaks, prompt injection, leakage and more.
- Author: Theo Marchetti
- Tags: Directory, #dir-open

Category

Open-source LLM vulnerability scanner

Website

[garak.ai](https://garak.ai/?ref=humanoidthreats.com)

Last verified

November 1, 2026

## What is garak?

garak, short for Generative AI Red-teaming and Assessment Kit, is an open-source LLM vulnerability scanner. It probes whether a language model or LLM-backed application can be made to fail through hallucination, data leakage, prompt injection, misinformation, toxicity or jailbreaks. It was created by Leon Derczynski in 2023, is released under the Apache 2.0 license and has been hosted on NVIDIA's GitHub since November 2024, maintained by NVIDIA and the community.

## Key capabilities

- **Probes** for DAN-style jailbreaks, encoding-based injection, invisible Unicode, cross-site scripting and automated attack generation.
- **Detectors** that judge whether an attack worked, with HTML and JSON reports.
- **Many targets,** including Hugging Face, OpenAI, AWS Bedrock, NVIDIA NIM, llama.cpp and generic REST endpoints.
- **Command-line interface** with dozens of plugins.
- **Built into NVIDIA's NeMo platform** for auditing deployed models.

## Why it matters for humanoid robot security

Editorial analysis: many humanoid makers build on open models or NVIDIA's robotics software, and garak gives them a free, scriptable way to test the language component for known failure modes. It tests models, not robots: it has no documented agent or physical-action features. That makes it a baseline check for the text and voice interfaces of a robot, best paired with agent-level and simulator-based testing.

## Where it fits in a humanoid robot security program

Analysis: garak fits at the AI model layer during design and testing. A robot team can point it at the language model behind a humanoid's voice assistant or task planner, including a self-hosted model exposed through a REST endpoint, and track results across model versions. It does not test the vision-language-action policy that drives motion directly, so it should be one input among several in a robot AI test plan.

## Who it is for

- Robot teams self-hosting open language models
- Researchers studying jailbreaks of embodied AI systems
- Security engineers who want a free baseline LLM scan

## Questions to ask garak

- Which probes are most relevant to a voice-controlled robot assistant?
- How do I connect garak to a self-hosted model on a robot or edge server?
- Can custom probes encode robot-specific forbidden commands?
- How should results be tracked over model updates?

## Similar listings

- [SROS2 (ROS 2 security)](https://www.humanoidthreats.com/directory/sros2/): Open-source robot security
- [MITRE ATLAS](https://www.humanoidthreats.com/directory/mitre-atlas/): AI threat framework
- [OWASP GenAI Security Project](https://www.humanoidthreats.com/directory/owasp-genai-security-project/): Open-source AI security guidance

## Related on Humanoid Threats

- [Can LLM-controlled robots be jailbroken?](https://www.humanoidthreats.com/can-llm-controlled-robots-be-jailbroken/)
- [Prompt injection against robot VLA models](https://www.humanoidthreats.com/prompt-injection-robot-vla-models/)
- [Browse the full Humanoid Robot Security Directory](https://www.humanoidthreats.com/directory/)

## Sources

- [garak website](https://garak.ai/?ref=humanoidthreats.com)
- [NVIDIA/garak on GitHub](https://github.com/NVIDIA/garak?ref=humanoidthreats.com)
- [NVIDIA NeMo docs: Vulnerability Scanning](https://docs.nvidia.com/nemo-platform/latest/documentation/vulnerability-scanning?ref=humanoidthreats.com)

*This listing is editorial and free; inclusion is not an endorsement. Work at garak?* [*Claim or update this listing*](mailto:info@humanoidthreats.com?subject=Update%20listing%3A%20garak-nvidia)*.*

Get the weekly Humanoid Threats Brief → [Subscribe](#/portal/signup)