Secure Use of Large Language Models for Cybersecurity Decision Support

Abstract

This thesis examines the secure use of large language models (LLMs) in cybersecurity decision-support contexts, with particular attention to reliability, privacy, and regulatory compliance. As LLMs are increasingly integrated into security workflows, concerns arise regarding their tendency to produce plausible but incorrect outputs, as well as their handling of sensitive data. These challenges are especially significant in environments governed by strict data protection regulations. The thesis evaluates the suitability of LLMs as assistive tools for cybersecurity analysts by analysing their performance characteristics, limitations in reasoning, and susceptibility to errors. In parallel, it investigates privacy risks associated with data processing, including issues related to data locality, leakage, and third-party model deployment. Based on this analysis, the thesis proposes a framework for the controlled and secure use of LLMs in cybersecurity settings. The framework emphasises human oversight, data minimisation, and alignment with existing governance models. The findings suggest that while LLMs can enhance efficiency and support analytical tasks, they should not be treated as authoritative decision-making systems. Instead, their value lies in augmenting human expertise within well-defined operational and regulatory boundaries.