Bridging the Blind Spots: A Holistic Risk Model for Secure Deployment of Large Language Models

Abstract

Academic research on securing Large Language Models (LLMs) in cybersecurity currently exists in silos. To address this fragmentation, this study develops the 'Holistic Deployment Risk Model' (HDRM) through a qualitative thematic synthesis of nine 'cornerstone' articles, selected through purposive sampling to ensure a representative cross-section of technical, ethical, and organizational perspectives, consolidating technical vulnerabilities, autonomous agentic risks, and adversarial misuse with organizational governance and human elements to identify critical 'blind spots'. The model clusters associated risks into five holistic, interdependent layers - Governance, Data and Privacy, Model Behavior, Operational Security, and Integrations and Infrastructure - while illustrating how vulnerabilities can propagate across these interdependent layers. To help demonstrate concrete applicability for regulatory readiness and real-world uses, components of the model are qualitatively mapped to current AI risk management frameworks: NIST AI RMF 1.0 and ISO/IEC 23894. While the model is currently theoretical and requires further investigation to confirm its efficacy, it offers organizations an actionable checklist to approach secure LLM deployment. It emphasizes the critical need for established governance before technical implementation and the importance of human-in-the-loop management for highrisk workflows. Ultimately, the work highlights key areas of 'ethical security' necessary to responsibly develop and manage AI systems, including the mitigation of probabilistic decisionmaking impacts such as bias, misinformation, and privacy violations.