Red Teaming LLMs With Garak

Red-Teaming LLMs with garak

Legal and Authorized-Use Notice: This skill is for authorized AI security testing and educational purposes only. Probe only models, API keys, and endpoints you own or have explicit written permission to test. Automated probing of third-party LLM APIs may violate their terms of service and consume billable tokens. Unauthorized probing of systems you do not control may be illegal.

Overview

garak (Generative AI Red-teaming and Assessment Kit) is an open-source LLM vulnerability scanner maintained by NVIDIA. It plays the role that a network vulnerability scanner like Nessus plays for hosts, but for large language models: it sends thousands of adversarial prompts ("probes") at a target model, captures the generations, and runs automated "detectors" over the responses to decide whether each attempt succeeded. Probe families cover prompt injection (promptinject, latentinjection), jailbreaks (dan), training-data and system-prompt leakage (leakreplay), malware generation (malwaregen), cross-site-scripting payload emission (xss), encoding-based bypasses (encoding), toxicity, and more. garak is described in the paper "garak: A Framework for Security Probing Large Language Models" (arXiv:2406.11036) and is distributed from the NVIDIA/garak GitHub repository.

The scanner is generator-agnostic. It can target Hugging Face models loaded locally, OpenAI-compatible APIs, AWS Bedrock, Replicate, Cohere, NIM endpoints, GGUF/llama.cpp models, and arbitrary REST endpoints via a JSON generator spec. After a run, garak emits a .report.jsonl line-delimited log of every attempt and detector verdict, a human-readable .report.html, a garak.log debug log, and a hit log of confirmed vulnerabilities. The terminal output prints a per-probe, per-detector pass/fail summary with a hit rate (for example dan.Dan_11_0 jailbreak: FAIL ok on 38/40), which is the primary artifact you interpret.

This skill maps to the MITRE ATLAS techniques AML.T0051 (LLM Prompt Injection) and AML.T0054 (LLM Jailbreak) because garak operationalizes both: it crafts prompt-injection and jailbreak inputs at scale and measures whether the target's guardrails hold. It supports the NIST AI RMF MEASURE-2.7 subcategory by providing repeatable, quantitative security/resilience measurement of a deployed AI system.

When to Use

• When you need a fast, repeatable baseline security assessment of an LLM before or after deployment.
• When validating that a guardrail, system prompt, or safety fine-tune actually reduces jailbreak and injection success rates (run before/after and compare hit rates).
• When producing evidence for an AI risk assessment or model card security section (NIST AI RMF MEASURE-2.7).
• When triaging which OWASP LLM Top 10 risks (LLM01 prompt injection, LLM02 sensitive information disclosure, LLM07 system prompt leakage) actually manifest in your model.
• When regression-testing an LLM endpoint in CI after model or prompt changes.

Prerequisites

• Python 3.10+ (3.12 recommended) and a virtual environment.
• Install garak from PyPI:

  python -m venv .venv && source .venv/bin/activate   # Windows: .venv\Scripts\activate
  python -m pip install -U garak
  garak --version

• For the bleeding-edge version:

  python -m pip install -U git+https://github.com/NVIDIA/garak.git@main

• An API key for the target if probing a hosted model (for example export OPENAI_API_KEY="sk-...").
• Written authorization to test the target, and awareness of token cost (probes generate thousands of calls).

Objectives

• Enumerate available probes and detectors and pick a scoped suite.
• Configure garak against a local Hugging Face model, an OpenAI-compatible API, and an arbitrary REST endpoint.
• Run prompt-injection, jailbreak, and leakage probe families and capture reports.
• Interpret the per-probe hit-rate output and the .report.jsonl.
• Re-run after applying a mitigation to demonstrate risk reduction.
• Produce a defensible findings artifact for an AI risk assessment.

MITRE ATT&CK Mapping

This skill uses MITRE ATLAS (the adversarial-ML companion to ATT&CK) technique IDs.

Workflow

Phase 1: Enumerate Probes and Detectors

1. List every probe garak ships so you can scope the run:

   garak --list_probes

2. List detectors (the modules that score whether a probe succeeded) and generators (target connectors):

   garak --list_detectors
   garak --list_generators

3. Read the probe taxonomy. Key families:
   • promptinject — PromptInject-framework direct injection.
   • latentinjection — instructions hidden in documents/encoded text (indirect injection).
   • dan — "Do Anything Now" and related jailbreaks (e.g. dan.Dan_11_0).
   • leakreplay — coax the model into reproducing memorized/training or hidden-prompt text.
   • encoding — base64/ROT13/etc. injection bypasses.
   • xss — emit cross-site-scripting payloads (markdown/HTML exfil).
   • malwaregen — request AV-evading or malicious code.

Phase 2: Probe a Local Hugging Face Model

1. Run a single jailbreak probe against a local model to validate setup:

   python -m garak --target_type huggingface --target_name gpt2 --probes dan.Dan_11_0

2. Run a fuller suite against a chat model:

   python -m garak \
     --target_type huggingface \
     --target_name meta-llama/Llama-3.2-1B-Instruct \
     --probes promptinject,dan,leakreplay \
     --report_prefix llama32_baseline