A model of ensuring LLM cybersecurity

The subject of study is a model for ensuring cybersecurity of Large Language Models (LLM). The goal of this study is to develop and analyze the components of the LLM cybersecurity model to improve its assessment accuracy and ensure the required security level. Tasks: the abstract structure of LLM systems should be suggested and analyzed; a common model of cybersecurity of LLM systems (LLMS) should be built; a cybersecurity model of LLM as a main component of LLMS should be developed; the elements of the developed cybersecurity model should be analyzed; potential case studies should be described and an example of risk criticality analysis for one of the threats of the LLM should be provided; the directions of future research should be substantiated on the identification, classification, criticality analysis, and collection of exploits to test the stability of LLM. The research results: the basic high-level architecture of LLMS, which consists of external sources, the LLM service, server functions, and storage environments, is developed; a common LLM cybersecurity model was built based on this architecture; the cybersecurity model was developed, which is an independent component of the overall cybersecurity model of LLMS and is based on a chain of the following elements: threat, vulnerability, attack, risks, and countermeasures; in addition, an analysis of the elements of the LLM cybersecurity model is conducted, and a sequence of countermeasures is proposed. Conclusions. This study determines that improving the cybersecurity of LLM is an important and urgent task, given the widespread use of these models in many areas of human life. The importance of developing an LLM cybersecurity model is that it is the baseline for all subsequent research. The practical significance of analyzing the model’s elements lies in using them to conduct experiments to simulate cyber attacks on LLM. The main contributions of this study are the LLM and LLMS cybersecurity models, the formalization of the results of these experiments, an assessment of the criticality level for cyber risks of the models, and the choice of countermeasures based on the coefficient of their effectiveness. In this case, ensuring an acceptable risk level for LLM is possible at a minimal cost. Areas for further research: definition and classification of exploits to test LLM security; methodology for collecting these exploits; analysis of the criticality of the damage they cause for various applications.