SyzMutateX: Optimizing Syzkaller Fuzzing with LLM-Driven Mutation and Adaptive Energy Allocation

Security vulnerabilities in the Linux kernel, a widely used core system component, pose a serious threat to the information infrastructure. Fuzzy testing is a key technique for mining vulnerabilities, and fuzzy tools like Syzkaller usually rely on random mutation strategies, which may lead to wasting computational resources on invalid mutations, and static random mutation weights may also lead to wasted resources. To address this problem, we provide SyzMutateX, which optimizes the mutation process for Linux kernel fuzzy testing by combining a Large Language Model (LLM) with an adaptive energy allocation technique. In order to provide the LLM with a rich context for generating efficient mutation seeds, SyzMutateX uses abstraction extraction techniques to extract contextual information such as file lists, function definitions, structure definitions, and dependencies from the Linux kernel source code. We present a mutation energy allocation algorithm (multi-armed slot machine) based on an upper confidence bound (UCB) with a sliding window, which dynamically selects mutation strategies by balancing exploration and exploitation while adapting to recent performance trends. In addition, our approach uses well-designed hints to perform five different mutation operations-inserting relevant system calls, eliminating redundant calls, changing parameters, splicing seeds, and performing pointer memory mutations-our approach integrates with 50% probability the LLM to generate target mutation seeds. Seed integrity is ensured by a final validation phase using Syzkaller's integrated syntax checker. Based on the experiments, SyzMutateX improved code coverage by 15.8% compared to the native Syzkaller. It also found nine different defects that have been reported and verified by Linux maintainers. These findings show how LLM and adaptive energy allocation can be combined to improve the efficiency and effectiveness of fuzz testing.