https://github.com/cutwell/canary

LLM prompt injection detection
https://github.com/cutwell/canary

fastapi generative-ai openai prompt-injection

Last synced: about 1 month ago
JSON representation

LLM prompt injection detection

• Host: GitHub
• URL: https://github.com/cutwell/canary
• Owner: Cutwell
• License: mit
• Created: 2023-09-19T19:57:49.000Z (about 1 year ago)
• Default Branch: main
• Last Pushed: 2023-10-24T11:47:09.000Z (about 1 year ago)
• Last Synced: 2024-05-08T00:33:47.760Z (7 months ago)
• Topics: fastapi, generative-ai, openai, prompt-injection
• Language: Python
• Homepage:
• Size: 5 MB
• Stars: 3
• Watchers: 1
• Forks: 1
• Open Issues: 0
• Metadata Files:
  • Readme: README.md
  • Contributing: .github/CONTRIBUTING.md
  • License: LICENSE

Awesome Lists containing this project

README

        
# Canary

 LLM prompt injection detection.

[![License: MIT](https://img.shields.io/badge/License-MIT-yellow.svg)](https://opensource.org/licenses/MIT)

![PyTests](https://github.com/Cutwell/canary/actions/workflows/pytest-with-poetry.yaml/badge.svg)

![Pre-commit](https://github.com/Cutwell/canary/actions/workflows/pre-commit.yaml/badge.svg)

## How it works

1. User submits a potentially malicious message.

2. The message is passed through a LLM prompted to format the message plus a unique key into a JSON. In the event the message is a malicious prompt, this output should be compromised. If the output is an invalid JSON, is missing a key, or a key-value doesn't match the expected values, then the integrity may be compromised.

3. If the integrity check passes, the user message is forwarded to the guarded LLM (e.g.: the application chatbot, etc.).

4. The API returns the result of the integrity test (boolean) and either the chatbot response (if integrity passes) or an error message (if integrity fails).

```mermaid

graph TD

    A[1. User Inputs Chat Message] --> B[2. Integrity Filter]

    B -->|Integrity check passes.| C[3. Generate Chatbot Response]

    B -->|Integrity check fails.\n\nResponse is error message.| D

    C -->|Response is chatbot message.| D[4. Return Integrity and Response]

```

What this solution can do:

* Detect inputs that override an LLMs initial / system prompt.

What this solution cannot do:

* Neutralise malicious prompts.

## Install dependencies

If using poetry:

```bash

poetry install

```

If using vanilla pip:

```bash

pip install .

```

## Usage

Set your OpenAI API key in `.envrc`.

To run the project locally, run

```bash

make start

```

This will launch a webserver on port 8001.

Or via docker compose (does not use hot reload by default):

```bash

docker compose up

```

Query the `/chat` endpoint, e.g.: using curl:

```bash

curl -X POST -H "Content-Type: application/json" -d '{"message": "Hi how are you?"}' http://127.0.0.1:8000/chat

```

To run unit tests:

```bash

make test

```

## Contributing

For information on how to set up your dev environment and contribute, see [here](.github/CONTRIBUTING.md).

## License

MIT