At the Wharton AI & Analytics Initiative Hackathon, our team reverse-engineered system instructions and crafted adversarial prompts to bypass RAG-based safeguards on a custom GPT chatbot. By analyzing model outputs and iteratively manipulating prompts, we exposed sensitive data and altered model behavior. The project earned 1st Place out of 60 teams for its demonstration of advanced prompt engineering and LLM security testing techniques.
At the Wharton AI & Analytics Initiative Hackathon, our team reverse-engineered system instructions and crafted adversarial prompts to bypass RAG-based safeguards on a custom GPT chatbot. By analyzing model outputs and iteratively manipulating prompts, we exposed sensitive data and altered model behavior. The project earned 1st Place out of 60 teams for its demonstration of advanced prompt engineering and LLM security testing techniques.
