Authors: Spencer Thompson
Mentors: Sayeed Sajal
Insitution: Utah Valley University
The widespread use of Large Language Models (LLMs) in various industries raises critical concerns about user data privacy and security. This research focuses on two key vulnerabilities: prompt attacks and unauthorized retrieval of sensitive training data. We employ a straightforward methodology to craft effective system messages that neutralize malicious queries in real-time, thereby mitigating prompt attacks. To prevent the unauthorized extraction of sensitive information, we build on the concept of strong system messages. We aim to identify a system message that minimizes computational overhead while maximizing effectiveness. Our results demonstrate that a strategically-crafted system message can guide an LLM's output in a manner that enhances data security without compromising computational efficiency.