Виявлення вразливостей за допомогою великих мовних моделей у програмних рішеннях, орієнтованих на беспілотні літальні апарати

This study considers binary software samples that operate and control unmanned aerial vehicles (UAVs). The task addressed is to detect vulnerabilities in UAV software given the absence of application source code. An improved method for automated vulnerability detection has been proposed, as well as a corresponding algorithm, a universal instruction template, and an architectural model for automated vulnerability search involving the capabilities of large language models (LLMs). Compared to the fuzzing method, the proposed method provides an average increase in accuracy to 94.7% while reducing the analysis time by 4 times. The method proposed for detecting UAV software vulnerabilities integrates binary analysis tools with the capabilities of logical inference and LLM pattern recognition. The corresponding algorithm for detecting UAV software vulnerabilities consists of processing stages, static analysis, logical inference using LLM, verification, correlation with known vulnerabilities, and reporting. The instruction template is independent of the features of the sample and tools and provides accurate logical conclusions. A new architectural communication model based on the Model-Context Protocol (MCP) provides universal interaction between LLM and decompilation tools. A comparative analysis of the method's applications for different implementations of cloud LLMs was carried out. Key advantages include the generation of detailed vulnerability reports, decreasing analysis time from hours to minutes through automation, as well as reducing the qualification requirements for reverse engineers who perform the analysis. The proposed solutions enable proactive security assessment of UAV software, as well as automated vulnerability detection