Methodology for fuzzing Linux kernel system calls using large language models

Authors: Teplyuk P. A., Yakunin A. G.

Annotation: A pressing issue in organizing Linux kernel fuzzing testing is creating system call specifications – special declarative descrip-tions that are subsequently used by a fuzzer to generate system call sequences. This is mostly manual work that requires deep knowledge, takes a lot of time, and does not exclude the error factor. Research is currently underway to automate the process of creating such specifications. The paper considers approaches to generate system call specifications KSG, SyzDescribe, and KernelGPT that have proven themselves in detecting unique kernel crashes during fuzz testing. A methodology to organize Linux kernel fuzzing testing is proposed, that includes a stage of automatic generation of system call specifications based on large language models (Large Language Model – LLM).

Keywords: llm, system call specification, syzkaller, fuzzing, linux kernel, operating system