Securing the Generative Frontier: A Systematic Analysis of Training Data Poisoning and Prompt Engineering Vulnerabilities in Large Language Models
Year: 2025
Type: article
Abstract: The rapid advancement and widespread adoption of generative AI technologies have introduced unprecedented cybersecurity challenges in data management and system security. This article presents a comprehensive analysis of security vulnerabilities in generative AI systems, with particular emphasis on training data poisoning attacks and prompt engineering exploits. Through systematic investigation of current security frameworks and emerging threat vectors, this article identifies critical vulnerabilities in training data management processes. It examines potential information leakage mechanisms in large language models. The article employs a multi-layered analytical approach to evaluate both intentional and unintentional security risks, ranging from malicious data manipulation to inadvertent exposure of sensitive information through model responses. Additionally, the article proposes a novel framework for threat detection and mitigation, incorporating adaptive security protocols and robust validation mechanisms for prompt engineering. The findings contribute to the growing cybersecurity literature by providing actionable insights for organizations implementing generative AI systems while establishing a foundation for future research in AI security standardization and self-healing machine learning models.
Author: Manojava Bharadwaj Bhagavathula
Institution: University of Pittsburgh
Language: English
Subfield: Artificial Intelligence
Field: Computer Science
Domain: Physical Sciences
Open Access status: diamond
OpenAlex ID: W4407302649