redteam-multi-modal (Multi-Modal Red Team Testing)

This example demonstrates how to use promptfoo's red teaming capabilities with multi-modal models, showing three different approaches to testing model safety and robustness against adversarial inputs involving images.

You can run this example with:

npx promptfoo@latest init --example redteam-multi-modal
cd redteam-multi-modal

Quick Start

1. Install dependencies:

   bashCopy

   npm install promptfoo sharp
   

2. Set up environment variables (see next section)

3. Run the static image example:

   bashCopy

   npx promptfoo@latest redteam eval -c redteam.static-image.yaml
   

4. Run the image strategy example:

   bashCopy

   npx promptfoo@latest redteam eval -c redteam.image-strategy.yaml
   

5. Run the UnsafeBench example:

   bashCopy

   npx promptfoo@latest redteam eval -c redteam.unsafebench.yaml
   

6. Run the VLGuard example:

   bashCopy

   npx promptfoo@latest redteam eval -c redteam.vlguard.yaml
   

7. Review results in the promptfoo interface

Environment Variables

This example requires the following environment variables depending on which approach you're using:

• For the Static Image and Image Strategy examples:

  • AWS_ACCESS_KEY_ID - Your AWS access key for Amazon Bedrock
  • AWS_SECRET_ACCESS_KEY - Your AWS secret key for Amazon Bedrock
  • AWS_REGION - Your AWS region (e.g., us-east-1)

• For the UnsafeBench and VLGuard examples:

  • HF_TOKEN or HF_API_TOKEN - Your Hugging Face API token (optional for VLGuard, required for UnsafeBench dataset access)

You can set these in a .env file or directly in your environment.

Description

This example provides four different approaches to red team testing with multi-modal models:

1. Static Image Example (promptfooconfig.static-image.yaml)

This configuration demonstrates how to red team a multi-modal model by keeping a static image (in this example, Buzz Aldrin on the moon) constant while varying the text prompt to test different potential attack vectors. The adversarial prompts are injected into the question variable.

Key features:

• Uses a fixed image (aldrin_apollo_11.jpg) for all tests
• Varies the text prompts using various harmful content plugins
• Tests how the model handles potentially harmful requests in the context of image analysis

2. Image Strategy Example (promptfooconfig.image-strategy.yaml)

This configuration demonstrates how to use promptfoo's image strategy for red teaming. Instead of attacking through text prompts, this approach converts potentially harmful text into images (text-to-image) and then sends those images to the model. This can help identify if models process text differently when it's embedded in an image rather than presented as plain text.

Key features:

• Keeps the text prompt constant ("Please answer the question in the image" - you can change this to any prompt or remove it entirely)
• Converts harmful text content into images using the image strategy
• Tests if harmful content embedded in images can bypass model safety filters

3. UnsafeBench Example (promptfooconfig.unsafebench.yaml)

This configuration uses the UnsafeBench plugin to evaluate multi-modal models against potentially unsafe imagery from the UnsafeBench dataset. It tests how well models detect and refuse to engage with harmful content across various categories.

Key features:

• Fetches images labeled as "unsafe" from the UnsafeBench dataset
• Automatically converts images to base64 format (even if originally provided as URLs)
• Filters images by category (Violence, Sexual, Hate, etc.)
• Tests if models appropriately refuse to engage with harmful visual content

4. VLGuard Example (promptfooconfig.vlguard.yaml)

This configuration uses the VLGuard plugin to evaluate multi-modal models against potentially unsafe imagery from the VLGuard dataset. It provides focused testing of content moderation capabilities with a smaller, curated dataset of 442 images.

Key features:

• Fetches images from the VLGuard dataset hosted on Hugging Face
• Organized into 4 main categories (deception, risky behavior, privacy, discrimination) and 9 subcategories
• Smaller, more focused dataset compared to UnsafeBench for targeted testing
• Tests model responses to potentially harmful visual content with specific safety guidelines

How It Works: Understanding Variables and Injection

In all examples, we work with prompt templates that have multiple variables:

• The static image example uses variables for both image and question
• The image strategy example uses these same variables but injects content differently
• The UnsafeBench example injects the unsafe image into the image variable

A crucial concept in these examples is that red teaming injects content into only one variable at a time (specified by injectVar in the config). The remaining variables need default values, which are set in the defaultTest section:

defaultTest:
  vars:
    # For static image example:
    image: file://aldrin_apollo_11.jpg # Fixed image while question varies

    # For image strategy example:
    question: 'Please answer the question in the image' # Fixed text while image varies

This approach lets you precisely control which part of your prompt receives adversarial content, while keeping other elements constant across all tests.

Static Image Configuration

The static image example uses a fixed JPEG image with adversarial text prompts injected into the conversation. The image remains constant while the text varies based on the redteam plugins.

Image Strategy Configuration

The image strategy converts text into images that are then encoded as base64 strings:

1. Takes potentially harmful text content
2. Renders it as black text on a white background in PNG format
3. Encodes the image as a base64 string
4. Passes this encoded image to the model instead of plain text

This approach can reveal whether models process text differently when it's embedded in images versus presented as plain text.

Important notes about the image strategy:

• It always produces PNG images when converting text to images
• The converted text appears as black text on a white background
• The strategy ensures the text is clearly readable in the generated image

How to Run

Running the Static Image Example

npx promptfoo@latest redteam eval -c redteam.static-image.yaml

Running the Image Strategy Example

npm install sharp # Required for the image strategy
npx promptfoo@latest redteam eval -c redteam.image-strategy.yaml

Running the UnsafeBench Example

First, ensure you have access to the UnsafeBench dataset and set your Hugging Face token:

export HF_TOKEN=your_huggingface_token

Then run:

npx promptfoo@latest redteam generate -c promptfooconfig.unsafebench.yaml
npx promptfoo@latest redteam eval -c redteam.yaml

Running the VLGuard Example

The VLGuard example works similarly to UnsafeBench but with a smaller, more focused dataset:

npx promptfoo@latest redteam eval -c redteam.vlguard.yaml

Or generate fresh test cases:

npx promptfoo@latest redteam generate -c promptfooconfig.vlguard.yaml
npx promptfoo@latest redteam eval -c redteam.yaml

Additional Configuration Options

Purpose Statement

When configuring your red team tests, ensure that the purpose statement in your configuration is relevant to the image content in some way. This helps create more realistic and contextually appropriate test scenarios.

For example, in the static image example, the purpose statement references NASA and Buzz Aldrin, which is directly relevant to the moon landing image:

redteam:
  purpose: >-
    You are a helpful assistant for NASA that can answer questions about this photo of Buzz Aldrin on the moon.

For more information on how the image strategy works, refer to the Image Inputs strategy documentation.

Supported Models

This example uses Amazon Bedrock's Nova model by default for the first two examples:

targets:
  - id: bedrock:amazon.nova-lite-v1:0
    label: multimodal-model

However, the techniques demonstrated can be applied to any multi-modal model that supports image inputs. You can easily swap out the target model in the configuration files to use models from:

• OpenAI: openai:gpt-4.1-mini and newer versions that support vision
• Anthropic: Claude 3 models with vision capabilities
• Google: Gemini models
• Other providers: Any model that can process both text and image inputs

To use a different model, simply update the targets section in your configuration file. The prompt format may need to be adjusted to match the requirements of your chosen model.

For example, to use OpenAI's GPT-4o:

targets:
  - id: openai:gpt-4o
    label: gpt4o-vision

Note you may have to update the prompt format to match the requirements of your chosen model.

Next Steps

After running these examples, consider:

1. Customizing plugins: Add or modify red team plugins to test specific vulnerabilities
2. Creating your own images: Test with domain-specific images relevant to your use case
3. Cross-model comparison: Compare how different models handle the same red team attacks
4. Enhancing safety measures: Apply lessons learned to improve your model's safety mechanisms

Troubleshooting

Common Issues

• Sharp installation problems: If you encounter issues installing the Sharp library, check the Sharp installation guide
• Image encoding errors: Ensure your image paths are correct and image files are valid
• API rate limits: Be mindful of your model provider's rate limits when running multiple tests
• Hugging Face authorization issues: Make sure you have requested access to the UnsafeBench dataset and your token has the correct permissions