michaelmainguy/saferprompt repository overview

⁠AI prompt injection detection.

Exposes API that uses local hugging face model to detect potentially malicious prompt injection. Should not be used as the sole line of defense, but as part of a comprehensive layered LLM security model.

Uses local only (doesn't call external APIs) protectai/deberta-v3-base-prompt-injection-v2 model...runs in embedded nodejs express server.

Source Code: https://github.com/mikemainguy/saferprompt⁠

The underlying engine can also be embedded as a library in a nodejs/browser app (though it requires 300m of storage for the local model....so browser might be a bad idea.

NPM Module: https://www.npmjs.com/package/saferprompt?activeTab=versions⁠

⁠Environment Variables

The container has two environment variables baked in via the Dockerfile:

There is also an optional variable not set by default:

⁠Overriding environment variables at runtime

Use -e flags to override any variable when running the container:

# Run on port 8080 with API key authentication enabled
docker run -p 8080:8080 \
  -e PORT=8080 \
  -e API_KEY=my-secret-key \
  saferprompt

Note: when you change PORT, update the -p mapping to match. The left side is the host port (your choice), the right side must match the container's PORT.

# Map host port 9090 to container port 8080
docker run -p 9090:8080 -e PORT=8080 saferprompt

⁠Using a .env file

You can pass a file of environment variables instead of individual -e flags:

# Create an env file
echo "API_KEY=my-secret-key" > .env.docker

# Pass it to docker run
docker run -p 3000:3000 --env-file .env.docker saferprompt

⁠Verifying the Container

# Health check (no API key)
curl -X POST http://localhost:3000/api/detect \
  -H "Content-Type: application/json" \
  -d '{"text": "hello"}'

# With API key
curl -X POST http://localhost:3000/api/detect \
  -H "Content-Type: application/json" \
  -H "x-api-key: my-secret-key" \
  -d '{"text": "hello"}'

Expected response:

{
  "label": "SAFE",
  "score": 0.9998,
  "isInjection": false,
  "ms": 42
}

You can also open http://localhost:3000⁠ in a browser to use the test UI.