9 garak(1) — exemplar reference page
10 garak(1) — exemplar reference page
garak — Generative AI Red-teaming & Assessment Kit (LLM vulnerability scanner)
| Layer | ATTACK |
| Category | A-2 — LLM Prompt Injection & Jailbreaking |
| Safety class | requires-target (refuses to run bare) |
| Licence | Apache-2.0 |
| Upstream | github.com/NVIDIA/garak |
10.1 NAME
garak — probes LLMs and dialog systems for hallucination, data leakage, prompt injection, misinformation, toxicity generation and jailbreaks: probes, detectors and buffs against a target model.
10.2 SYNOPSIS
garak --target_type TYPE --target_name NAME [-S SPEC] [OPTIONS]
garak --list_probes [-v] [--spec SPEC]
garak --plugin_info TYPE.PLUGIN.CLASS
garak --version10.3 DESCRIPTION
garak scans a target model with static, dynamic and adaptive probes, scores outputs with detectors, and modifies probe behaviour with buffs. All runs require a target generator — consistent with its requires-target safety class; the tool does not self-fire.
10.4 EXAMPLES
garak --target_type ollama --target_name llama3.2:latest --spec tag:owasp:llm01
garak --target_type openai --target_name gpt-4o-mini --spec probes.dan
garak --list_probes -v10.5 SEE ALSO
Sister A-2 tools in the same category; the A-2 category page; the packaging record (Kailash OS #81); the full sample this exemplar condenses lives in the kailash-website#30 review thread.
Version-lock note. Generated against the Kailash pin of the manifest row (pkgs/garak/default.nix + nvfetcher source), verified 2026-10-14; smoke status: recorded on the packaging issue (the darwin lane box has no Linux build surface — the provenance is stated honestly rather than implying a verified run). If the lock moves garak and this note no longer matches, the page is stale — regeneration is required in the same change that moves the version.
This page is the template acceptance exemplar: every generated reference page must render in this shape (the full template and the generation contract are ADR-0001 in this repository).