Garak

Garak is a command-line tool that scans large language models for vulnerabilities and failure modes.

Garak is a generative AI red-teaming and assessment kit that checks if a large language model can be made to fail in unwanted ways. It combines static, dynamic, and adaptive probes to explore vulnerabilities like hallucination, data leakage, prompt injection, misinformation, toxicity generation, and jailbreaks. The tool is designed for developers and security practitioners who need to assess model safety. It operates as a command-line interface and supports various platforms including Hugging Face, OpenAI, Replicate, Cohere, Groq, AWS Bedrock, and REST endpoints. Garak evaluates model responses using specific detectors and logs detailed results in JSONL format for analysis.

Garak at a glance

Best for: Software engineer
Runs on: Linux, macOS, CLI
Pricing: Free
Released: May 2023
Category: AI coding tool
Website: github.com/NVIDIA/garak

Key features

Probe for hallucinations and data leakage
Test for prompt injection and jailbreaks
Run static, dynamic, and adaptive probes
Evaluate model responses with detectors
Log run details in JSONL format
Support multiple model generators
Use command-line interface syntax

Use it for

Red-team generative AI models
Scan large language models for vulnerabilities
Test language models for toxicity
Assess model resistance to jailbreaks

Alternatives to Garak

Genkit — Provides unified APIs and local developer tools for AI application workflows.
Giskard — It offers automated vulnerability scanning and behavioral testing for LLM agents.
jevgrep — Reduces coding agent costs by providing targeted, relevant repository context.
Tusk — Generates tests from production traffic to catch regressions in PRs.
Kora Doctor — Surfaces waste across tool usage, context growth, and repeated inference without an account.
PentAGI — It autonomously manages penetration testing workflows within a secure Docker sandbox.
Mantis — Automates vulnerability discovery and patching using AI agents.
ARTEX — Runs autonomous multi-agent penetration testing with a dual-graph architecture.
Opik — It offers comprehensive tracing, automated evaluations, and open-source flexibility.
LLM Gauze — Remediates local language model errors and fixes malformed tool calls automatically.
RTK — Reduces AI context output noise by 56 to 90 percent with zero configuration.
Marmot — Consolidates multiple AI and data provider interfaces into one CLI.

All Garak alternatives