Hack - AI - mBank chat hacker

Whole documentation is https://Michall00.github.io/hack-ai-prompt-buddies/

Project Overview

mBank Chat Hacker is a project designed to test the robustness, security, and accuracy of mBank's chatbot system. The primary goal is to simulate real-world and edge-case scenarios to evaluate the chatbot's behavior, identify vulnerabilities, and improve its performance. The system leverages advanced AI models to generate prompts, analyze responses, and simulate user interactions.

Key Features:

• Dynamic Prompt Generation: The system uses a modular approach to generate prompts tailored to specific categories, such as hallucinations, misclassifications, and unauthorized behavior.
• Automated Chat Interaction: The chatbot is tested using Playwright to simulate real user interactions in a browser environment.
• Error Simulation: Tools are provided to simulate incorrect transactions, currency conversions, and other edge cases.
• Evaluation Framework: A Streamlit-based application allows for manual evaluation of chatbot responses, categorization of errors, and feedback collection.
• Logging and Analysis: All interactions are logged for further analysis and debugging.
• Multi-Agent Strategy with Wolf Selector: A unique system that dynamically selects between cooperative and adversarial agents to test the chatbot's resilience.

---

Wolf Selector: Multi-Agent Strategy

The Wolf Selector is a core component of the system, designed as a strategic decision-maker. It operates by analyzing the history of interactions with the chatbot and dynamically deciding which agent—good or bad—should generate the next prompt. This approach allows for a sophisticated simulation of both cooperative and adversarial user behaviors.

How It Works:

1. Good Agent:
   • Behaves like a typical user: polite, cooperative, and non-suspicious.
   • Focuses on building trust and calming the chatbot to avoid triggering defensive mechanisms.
2. Bad Agent:
   • Acts as an adversarial user: manipulative, provocative, and designed to test the chatbot's limits.
   • Attempts to exploit vulnerabilities, provoke errors, or bypass security mechanisms.
3. Wolf Selector:
   • Acts as a strategist that evaluates the conversation history and decides which agent should take control.
   • Uses a system prompt to analyze the context and determine whether to escalate (use the bad agent) or de-escalate (use the good agent).
   • Ensures a balanced approach by alternating between cooperative and adversarial behaviors to avoid detection by the chatbot.

Implementation Details

Architecture

The project is built using a modular architecture to ensure scalability and maintainability. Key components include:

1. Prompt Generation:

   • The PromptGenerator class dynamically generates prompts based on predefined categories (e.g., hallucinations, misclassifications).
   • Prompts are enriched with examples and guidelines to simulate realistic user interactions.

2. Chat Interaction:

   • The run function in main.py uses Playwright to automate interactions with the mBank chatbot.
   • The WolfSelector class dynamically selects between "good" and "bad" prompt generators to simulate different user behaviors.

3. Error Simulation:

   • Tools in tools.py provide functionality to simulate fake transactions, incorrect balances, and currency conversion errors.

4. Evaluation:

   • A Streamlit-based app (app.py) allows for manual evaluation of chatbot responses, categorization of errors, and feedback collection.

5. Logging:

   • All interactions are logged using logging_utils for debugging and analysis.

Approach

The project adopts a test-driven approach to evaluate the chatbot's performance. Key aspects include:

• Edge Case Testing: Simulating scenarios that push the chatbot to its limits, such as handling ambiguous queries or unauthorized requests.
• Dynamic Behavior: Using the WolfSelector to alternate between cooperative and adversarial user behaviors.
• Realistic Simulations: Generating prompts that mimic real-world user interactions, including emotional and manipulative language.

---

Getting Started

Clone repo

git clone https://github.com/Michall00/hack-ai-prompt-buddies.git

Prerequisites

Before starting with the project, make sure you have installed all the required dependencies. You can do this by running the following command:

make create_environment

make requirements

Set up env

Create .env file in main directory and fill

TOGETHER_API_KEY = ...
LOGIN = ...
PASSWORD = ... 

Have fun

Automate chat with mBank

make run

Run evaluation app

make run_evaluation

Team

Natalia Pieczko
Daniel Machniak
Krzystof Gólcz
Mateusz Ostaszewski
Michał Sadowski