Chenyuan Yang, Zijie Zhao, Lingming Zhang

• Read: 21 Jun 2025
• Published: 30 Mar 2025

ASPLOS ‘25: Proceedings of the 30th ACM International Conference on Architectural Support for Programming Languages and Operating Systems, Volume 2 Pages 560 - 573

https://doi.org/10.1145/3676641.3716022

---

Q&A (link)

What are the motivations for this work?

• See Introduction.
• Key insight is that LLMs have seen massive kernel code, documentation, and use cases during pre-training, and thus can automatically distill the necessary information for making valid syscalls.

What is the proposed solution?

• See Introduction.
• KernelGPT leverages an iterative approach to automatically infer the specifications, and further debug and repair them based on the validation feedback.

What is the work’s evaluation of the proposed solution?

• See Section 5.

What is your analysis of the identified problem, idea and evaluation?

NONE

What are the contributions?

• See Introduction.
• Proposed the first automated approach to leveraging the potential of LLMs for kernel fuzzing.
• Implemented KernelGPT to infer syscall specifications with a novel iterative strategy.
• Detected 24 previously unknown bugs.

What are future directions for this research?

NONE

What questions are you left with?

NONE

What is your take-away message from this paper?

NONE