PROBLEM ISSUES IN USING LARGE LANGUAGE MODELS FOR DECOMPILATION OF MACHINE CODE WITH VULNERABILITIES

This paper examines the problem of software vulnerabilities in the absence of source code. One way to counter them is by decompilation the machine (executable) code of programs. The paper considers the application of a relatively new technology, large language models, to the task of restoring pseudo-source code suitable for detecting and eliminating vulnerabilities. The paper identifies problematic issues in the subject area, such as the incompleteness of the dataset for rare processor architectures, the lack of a guarantee that the obtained source code is identical to the specified machine code, the sanitization of the recovered source code by fixing vulnerabilities, hallucinations in the code, and the difficulty of restoring obfuscated (including optimized) code. To substantiate and demonstrate the essence of each problematic issue, a practical example of decompilation assembly code functions using the widespread large language model DeepSeek-V3.2 is provided. The negative impact of these problematic issues on the final neutralization of vulnerabilities is also indicated.