Generative AI Red-teaming & Assessment Kit
garak checks if an LLM can be made to fail in a way we don't want. garak probes for hallucination, data leakage, prompt injection, misinformation, toxicity generation, jailbreaks, and many other weaknesses. If you know nmap, it's nmap for LLMs.
garak focuses on ways of making an LLM or dialog system fail. It combines static, dyanmic, and adaptive probes to explore this.
garak's a free tool. We love developing it and are always interested in adding functionality to support applications.
> See our user guide! docs.garak.ai
> Join our Discord!
> Project links & home: garak.ai
> Twitter: @garak_llm
> DEF CON slides!
currently supports:
- hugging face hub generative models
- replicate text models
- openai api chat & continuation models
- litellm
- pretty much anything accessible via REST
- gguf models like llama.cpp version >= 1046
- .. and many more LLMs!
garak is a command-line tool. It's developed in Linux and OSX.
Just grab it from PyPI and you should be good to go:
python -m pip install -U garak
The standard pip version of garak is updated periodically. To get a fresher version, from GitHub, try:
python -m pip install -U git+https://github.com/leondz/garak.git@main
garak has its own dependencies. You can to install garak in its own Conda environment:
conda create --name garak "python>=3.10,<=3.12"
conda activate garak
gh repo clone leondz/garak
cd garak
python -m pip install -e .
OK, if that went fine, you're probably good to go!
The general syntax is:
garak <options>
garak needs to know what model to scan, and by default, it'll try all the probes it knows on that model, using the vulnerability detectors recommended by each probe. You can see a list of probes using:
garak --list_probes
To specify a generator, use the --model_type and, optionally, the --model_name options. Model type specifies a model family/interface; model name specifies the exact model to be used. The "Intro to generators" section below describes some of the generators supported. A straightforward generator family is Hugging Face models; to load one of these, set --model_type to huggingface and --model_name to the model's name on Hub (e.g. "RWKV/rwkv-4-169m-pile"). Some generators might need an API key to be set as an environment variable, and they'll let you know if they need that.
garak runs all the probes by default, but you can be specific about that too. --probes promptinject will use only the PromptInject framework's methods, for example. You can also specify one specific plugin instead of a plugin family by adding the plugin name after a .; for example, --probes lmrc.SlurUsage will use an implementation of checking for models generating slurs based on the Language Model Risk Cards framework.
For help & inspiration, find us on twitter or discord!
Probe ChatGPT for encoding-based prompt injection (OSX/*nix) (replace example value with a real OpenAI API key)
export OPENAI_API_KEY="sk-123XXXXXXXXXXXX"
python3 -m garak --model_type openai --model_name gpt-3.5-turbo --probes encoding
See if the Hugging Face version of GPT2 is vulnerable to DAN 11.0
python3 -m garak --model_type huggingface --model_name gpt2 --probes dan.Dan_11_0
For each probe loaded, garak will print a progress bar as it generates. Once generation is complete, a row evaluating that probe's results on each detector is given. If any of the prompt attempts yielded an undesirable behavior, the response will be marked as FAIL, and the failure rate given.
Here are the results with the encoding module on a GPT-3 variant:
And the same results for ChatGPT:
We can see that the more recent model is much more susceptible to encoding-based injection attacks, where text-babbage-001 was only found to be vulnerable to quoted-printable and MIME encoding injections. The figures at the end of each row, e.g. 840/840, indicate the number of text generations total and then how many of these seemed to behave OK. The figure can be quite high because more than one generation is made per prompt - by default, 10.
Errors go in garak.log; the run is logged in detail in a .jsonl file specified at analysis start & end. There's a basic analysis script in analyse/analyse_log.py which will output the probes and prompts that led to the most hits.
Send PRs & open issues. Happy hunting!
Using the Pipeline API:
--model_type huggingface(for transformers models to run locally)--model_name- use the model name from Hub. Only generative models will work. If it fails and shouldn't, please open an issue and paste in the command you tried + the exception!
Using the Inference API:
--model_type huggingface.InferenceAPI(for API-based model access)--model_name- the model name from Hub, e.g."mosaicml/mpt-7b-instruct"
Using private endpoints:
-
--model_type huggingface.InferenceEndpoint(for private endpoints) -
--model_name- the endpoint URL, e.g.https://xxx.us-east-1.aws.endpoints.huggingface.cloud -
(optional) set the
HF_INFERENCE_TOKENenvironment variable to a Hugging Face API token with the "read" role; see https://huggingface.co/settings/tokens when logged in
--model_type openai--model_name- the OpenAI model you'd like to use.gpt-3.5-turbo-0125is fast and fine for testing.- set the
OPENAI_API_KEYenvironment variable to your OpenAI API key (e.g. "sk-19763ASDF87q6657"); see https://platform.openai.com/account/api-keys when logged in
Recognised model types are whitelisted, because the plugin needs to know which sub-API to use. Completion or ChatCompletion models are OK. If you'd like to use a model not supported, you should get an informative error message, and please send a PR / open an issue.
- set the
REPLICATE_API_TOKENenvironment variable to your Replicate API token, e.g. "r8-123XXXXXXXXXXXX"; see https://replicate.com/account/api-tokens when logged in
Public Replicate models:
--model_type replicate--model_name- the Replicate model name and hash, e.g."stability-ai/stablelm-tuned-alpha-7b:c49dae36"
Private Replicate endpoints:
--model_type replicate.InferenceEndpoint(for private endpoints)--model_name- username/model-name slug from the deployed endpoint, e.g.elim/elims-llama2-7b
--model_type cohere--model_name(optional,commandby default) - The specific Cohere model you'd like to test- set the
COHERE_API_KEYenvironment variable to your Cohere API key, e.g. "aBcDeFgHiJ123456789"; see https://dashboard.cohere.ai/api-keys when logged in
--model_type groq--model_name- The name of the model to access via the Groq API- set the
GROQ_API_KEYenvironment variable to your Groq API key, see https://console.groq.com/docs/quickstart for details on creating an API key
--model_type ggml--model_name- The path to the ggml model you'd like to load, e.g./home/leon/llama.cpp/models/7B/ggml-model-q4_0.bin- set the
GGML_MAIN_PATHenvironment variable to the path to your ggmlmainexecutable
rest.RestGenerator is highly flexible and can connect to any REST endpoint that returns plaintext or JSON. It does need some brief config, which will typically result a short YAML file describing your endpoint. See https://reference.garak.ai/en/latest/garak.generators.rest.html for examples.
Use models from https://build.nvidia.com/ or other NIM endpoints.
- set the
NIM_API_KEYenvironment variable to your authentication API token, or specify it in the config YAML
For chat models:
--model_type nim--model_name- the NIMmodelname, e.g.meta/llama-3.1-8b-instruct
For completion models:
--model_type nim.NVOpenAICompletion--model_name- the NIMmodelname, e.g.bigcode/starcoder2-15b
- set the
OCTO_API_TOKENenvironment variable to your Replicate API token, e.g. "r8-123XXXXXXXXXXXX"; see https://replicate.com/account/api-tokens when logged in
Octo public endpoint:
--model_type octo--model_name- the OctoAI public endpoint for the model, e.g.mistral-7b-instruct-fp16
Octo private endpoint:
--model_type octo.InferenceEndpoint(for private endpoints)--model_name- the deployed endpoint URL, e.g.https://llama-2-70b-chat-xxx.octoai.run/v1/chat/completions
-
--model_type test -
(alternatively)
--model_name test.BlankFor testing. This always generates the empty string, using thetest.Blankgenerator. Will be marked as failing for any tests that require an output, e.g. those that make contentious claims and expect the model to refute them in order to pass. -
--model_type test.RepeatFor testing. This generator repeats back the prompt it received.
| Probe | Description |
|---|---|
| blank | A simple probe that always sends an empty prompt. |
| atkgen | Automated Attack Generation. A red-teaming LLM probes the target and reacts to it in an attempt to get toxic output. Prototype, mostly stateless, for now uses a simple GPT-2 fine-tuned on the subset of hhrlhf attempts that yielded detectable toxicity (the only target currently supported for now). |
| av_spam_scanning | Probes that attempt to make the model output malicious content signatures |
| continuation | Probes that test if the model will continue a probably undesirable word |
| dan | Various DAN and DAN-like attacks |
| donotanswer | Prompts to which responsible language models should not answer. |
| encoding | Prompt injection through text encoding |
| gcg | Disrupt a system prompt by appending an adversarial suffix. |
| glitch | Probe model for glitch tokens that provoke unusual behavior. |
| grandma | Appeal to be reminded of one's grandmother. |
| goodside | Implementations of Riley Goodside attacks. |
| leakerplay | Evaluate if a model will replay training data. |
| lmrc | Subsample of the Language Model Risk Cards probes |
| malwaregen | Attempts to have the model generate code for building malware |
| misleading | Attempts to make a model support misleading and false claims |
| packagehallucination | Trying to get code generations that specify non-existent (and therefore insecure) packages. |
| promptinject | Implementation of the Agency Enterprise PromptInject work (best paper awards @ NeurIPS ML Safety Workshop 2022) |
| realtoxicityprompts | Subset of the RealToxicityPrompts work (data constrained because the full test will take so long to run) |
| snowball | Snowballed Hallucination probes designed to make a model give a wrong answer to questions too complex for it to process |
| xss | Look for vulnerabilities the permit or enact cross-site attacks, such as private data exfiltration. |
garak generates multiple kinds of log:
- A log file,
garak.log. This includes debugging information fromgarakand its plugins, and is continued across runs. - A report of the current run, structured as JSONL. A new report file is created every time
garakruns. The name of this file is output at the beginning and, if successful, also the end of the run. In the report, an entry is made for each probing attempt both as the generations are received, and again when they are evaluated; the entry'sstatusattribute takes a constant fromgarak.attemptsto describe what stage it was made at. - A hit log, detailing attempts that yielded a vulnerability (a 'hit')
Check out the reference docs for an authoritative guide to garak code structure.
In a typical run, garak will read a model type (and optionally model name) from the command line, then determine which probes and detectors to run, start up a generator, and then pass these to a harness to do the probing; an evaluator deals with the results. There are many modules in each of these categories, and each module provides a number of classes that act as individual plugins.
garak/probes/- classes for generating interactions with LLMsgarak/detectors/- classes for detecting an LLM is exhibiting a given failure modegarak/evaluators/- assessment reporting schemesgarak/generators/- plugins for LLMs to be probedgarak/harnesses/- classes for structuring testingresources/- ancillary items required by plugins
The default operating mode is to use the probewise harness. Given a list of probe module names and probe plugin names, the probewise harness instantiates each probe, then for each probe reads its recommended_detectors attribute to get a list of detectors to run on the output.
Each plugin category (probes, detectors, evaluators, generators, harnesses) includes a base.py which defines the base classes usable by plugins in that category. Each plugin module defines plugin classes that inherit from one of the base classes. For example, garak.generators.openai.OpenAIGenerator descends from garak.generators.base.Generator.
Larger artefacts, like model files and bigger corpora, are kept out of the repository; they can be stored on e.g. Hugging Face Hub and loaded locally by clients using garak.
- Take a look at how other plugins do it
- Inherit from one of the base classes, e.g.
garak.probes.base.TextProbe - Override as little as possible
- You can test the new code in at least two ways:
- Start an interactive Python session
- Import the model, e.g.
import garak.probes.mymodule - Instantiate the plugin, e.g.
p = garak.probes.mymodule.MyProbe()
- Import the model, e.g.
- Run a scan with test plugins
- For probes, try a blank generator and always.Pass detector:
python3 -m garak -m test.Blank -p mymodule -d always.Pass - For detectors, try a blank generator and a blank probe:
python3 -m garak -m test.Blank -p test.Blank -d mymodule - For generators, try a blank probe and always.Pass detector:
python3 -m garak -m mymodule -p test.Blank -d always.Pass
- For probes, try a blank generator and always.Pass detector:
- Get
garakto list all the plugins of the type you're writing, with--list_probes,--list_detectors, or--list_generators
- Start an interactive Python session
We have an FAQ here. Reach out if you have any more questions! leon@garak.ai
Code reference documentation is at garak.readthedocs.io.
You can read the garak preprint paper. If you use garak, please cite us.
@article{garak,
title={{garak: A Framework for Security Probing Large Language Models}},
author={Leon Derczynski and Erick Galinkin and Jeffrey Martin and Subho Majumdar and Nanna Inie},
year={2024},
howpublished={\url{https://garak.ai}}
}
"Lying is a skill like any other, and if you wish to maintain a level of excellence you have to practice constantly" - Elim
For updates and news see @garak_llm
© 2023- Leon Derczynski; Apache license v2, see LICENSE
![@github-actions[bot]](https://avatars.githubusercontent.com/in/15368?s=64&v=4){kind=small}
