A System-Level Security Analysis and Roadmap for Large Language Model-Based Systems
Kitapta bölümler > Diğer · 2026
Yazarlar: Selin Şaşal, Özgü CAN
Advances in Transdisciplinary Engineering, Diğer, IOS Press, B, Uluslararası, 2026, ,
DOI: 10.3233/atde260482
Özet
The evolution of Large Language Models (LLMs) from passive text generators to centralized reasoning components in data retrieval and agent-based systems has revealed security challenges that extend beyond traditional model robustness concepts. Current assessment approaches, focusing primarily on static input-output behavior, are insufficient to capture vulnerabilities arising from system orchestration, external data retrieval, and autonomous action execution. This work presents a system-level security analysis of integrated LLM architectures, focusing on the interaction between models, orchestration layers, data retrieval mechanisms, and execution interfaces. For this purpose, the study examines these interactions to characterize an extended attack surface that facilitates threats such as indirect prompt injection, data poisoning through data retrieval processes, and scenarios in which models are manipulated to execute privileged or insecure actions. Therefore, a structured threat model is defined to address these risks, and a unified taxonomy of vulnerabilities relevant to contemporary LLM deployments is derived. The analysis further introduces an assessment framework centered on context integrity, execution security, and error propagation across system components, and provides practical engineering recommendations for secure deployment and operation. Consequently, this study addresses LLM safety as a system-level engineering challenge and provides a foundation for evaluating and enhancing LLM-based systems beyond isolated model behavior.