Prompt Injection, Data Leakage, and Safety Defense in Multimodal LLM-Integrated Systems

Abstract

The integration of multimodal large language models (LLMs) into socio-technical infrastructures has introduced unprecedented capabilities alongside novel security vulnerabilities. This paper examines the systemic risks posed by prompt injection attacks, data leakage pathways, and the corresponding defensive architectures required to maintain safety and robustness in deployed systems. Unlike prior work that isolates individual attack vectors, we adopt a holistic lens that considers the entire lifecycle of multimodal LLM-integrated systems, from model training and fine-tuning to runtime orchestration and governance. We argue that prompt injection exploits the inherent ambiguity between instruction and data in LLM interfaces, and that multimodal inputs exacerbate this ambiguity by adding heterogeneous encodings such as images, audio, and video. Data leakage, often a consequence of prompt injection or inadequate output filtering, raises critical concerns about privacy, fairness, and regulatory compliance. We propose a layered defense framework that combines input sanitization, context-aware output validation, and architecture-level isolation mechanisms, while acknowledging the fundamental trade-offs between security, usability, and computational cost. Through cross-domain analysis and case illustrations, we demonstrate that no single defense suffices; rather, a governance-oriented approach integrating technical controls with policy mechanisms is essential. Our discussion extends to sustainability implications, deployment challenges in resource-constrained environments, and the ethical imperative of fairness in defense design. We conclude with forward-looking perspectives on adversarial robustness, adaptive threat models, and the need for interdisciplinary collaboration in shaping the future of safe multimodal AI.