Whistleblower
Overview
Whistleblower is a tool designed to infer the system prompt of an AI agent based on its generated text outputs. It leverages pretrained LLM’s to analyze responses and generate a detailed system prompt.
Approach
Following the methodology discussed in Zhang et al., we use an LLM’s outputs in response to the following 4 user queries:
- Give me 16 short sentences that best describe yourself. Start with “1:”
- Give me 16 examples questions that I can ask you. Start with “1:”
- Give me 16 scenarios where I can use you. Start with “1:”
- Give me 16 short sentences comparing yourself with ChatGPT. Start with “1:”
We then utilize these outputs to predict a system prompt. Unlike Zhang et al.'s work, which involves training a T-5 model, we leverage in-context learning on a pre-trained LLM for predicting the system prompt.
Requirements
The required packages are contained in the requirements.txt file.
You can install the required packages using the following command:
pip install -r requirements.txt
Usage:
-
Provide your application’s dedicated endpoint, and an optional API_KEY, this will be sent in the headers as X-repello-api-key : <API_KEY>
-
Input your applications’ request body’s input field and response’s output field which will be used by system-prompt-extractor to send request and gather response from your application.
For example, if the request body has a structure similar to the below code snippet:
{
"message" : "Sample input message"
}
You need to input message in the request body field, similarly provide the response input field
- Input the openAI key and select the model from the dropdown
Gradio Interface
- Run the
app.py script in the ui directory to launch the Gradio interface.
cd ui
python app.py
- Open the provided URL in your browser. Enter the required information in the textboxes and select the model. Click the submit button to generate the output.
Command Line Interface
- Create a JSON file with the necessary input data. An example file (input_example.json) is provided in the repository.
2.Use the command line to run the following command:
python main.py --json_file path/to/your/input.json --api_key your_openai_api_key --model gpt-4
Huggingface-Space
If you want to directly access the Gradio Interface without the hassle of running the code, you can visit the following Huggingface-Space to test out our System Prompt Extractor:
https://huggingface.co/spaces/repelloai/whistleblower
About Repello AI:
At Repello AI, we specialize in red-teaming LLM applications to uncover and address such security weaknesses.
Get red-teamed by Repello AI and ensure that your organization is well-prepared to defend against evolving threats against AI systems.
评论