Google Introduces Sashiko: AI-Powered Code Review System for Linux Kernel Catches Bugs Humans Miss
Key Takeaways
- ▸Sashiko detected 53% of bugs missed by human reviewers in a sample of 1,000 recent Linux kernel issues
- ▸The system uses Google's Gemini 3.1 Pro LLM but is designed to work with other LLM providers like Claude
- ▸Unlike controversial AI code submission tools, Sashiko functions as a code review assistant to reduce maintainer burden
Summary
Google engineer Roman Gushchin has announced Sashiko, an AI-powered code review system designed to assist with Linux kernel development. Built in Rust and powered by Google's Gemini 3.1 Pro LLM, Sashiko analyzes patches submitted to the Linux Kernel Mailing List and provides feedback to maintainers and developers. According to measurements, the system identified 53% of bugs from a sample of 1,000 recent upstream issues—bugs that were entirely missed by human reviewers.
Sashiko works by ingesting patches from mailing lists, analyzing them with an LLM, and providing high-quality review feedback. The system has been tested internally at Google and now belongs to the Linux Foundation. While the false positive rate is estimated between 15-20%, the authors note that much of this falls into a "gray zone" of borderline issues. The tool is configured to work with multiple LLM providers including Gemini Pro 3.1 and Claude, though code and data are sent to external LLM services—a transparency the authors acknowledge upfront.
The initiative addresses growing pressure on Linux kernel maintainers who face overwhelming code review workloads. Unlike controversial AI code submission tools that have drawn criticism from the open source community, Sashiko positions itself as a supportive review assistant rather than a code contributor. Google is currently covering operational costs for running Sashiko on the Linux Kernel Mailing List.
- The tool transparently sends code and data to external LLM providers; Google covers operational costs for the Linux kernel project
- False positive rate is estimated at 15-20%, with many borderline cases falling into a gray zone
Editorial Opinion
Sashiko represents a pragmatic application of AI to open source development that may prove more palatable than automated code submission. By positioning AI as a reviewer's assistant rather than a contributor, Google sidesteps much of the philosophical objection plaguing other AI tools in FOSS communities. However, the reliance on external LLM providers raises legitimate privacy concerns for kernel code—even if transparently disclosed—and the 53% detection rate, while impressive for missed issues, suggests significant room for improvement before such tools become truly indispensable.
