An AI bug detection system called 'Sashiko' has been developed by a Google employee. Named after the Japanese embroidery technique 'sashiko,' it has been successfully detecting previously undiscovered bugs in the Linux kernel.
Google employee
Sashiko
https://sashiko.dev/
Sashiko is named after ' sashiko ,' a traditional Japanese embroidery technique. Sashiko is described as 'a technique that has decorative meaning as well as mere reinforcement (patching),' making it a perfect fit for the system's mission of 'strengthening the Linux kernel with automated patch review.'
Sashiko is a system that scans all patch submission emails sent to the Linux kernel project mailing list and analyzes them for bugs using Gemini 3.1 Pro. The Sashiko webpage archives the scanned emails and their analysis results.
Emails with 'Reviewed' in the Status column are those that have been patch-analyzed, and the 'Findings' column displays the number of issues with severity levels 'Critical,' 'High,' 'Medium,' and 'Low' from left to right. For example, the patch submitted in the email enclosed in the red box contains one issue with severity level 'Critical' and two issues with severity level 'High.'
Clicking on each email will open a details screen. The details screen displays the email's thread structure, patch contents, and AI review results.
The AI review results are displayed inline within the patch code.
Clicking 'View Raw Log' at the bottom of the screen will display the input and output data for the AI.
This is the log of the interaction with the AI. As you can see from the scroll bar on the far right, a very long exchange took place.
Sashiko was able to automatically detect 53% of the most recent 1000 problems. 100% of the problems Sashiko found had been overlooked by humans. As mentioned above, running Sashiko involves lengthy interactions with the AI and consumes a large amount of tokens, but Google provides the AI for running Sashiko in the Linux kernel project free of charge.
As of the time of writing, Sashiko is managed under the Linux Foundation. The source code is publicly available at the following link, and anyone can analyze their own code by self-hosting Sashiko and registering an AI API key. Sashiko has been tested with Gemini Pro 3.1, but it also supports Claude, and developer Gushchin states that 'theoretically it will work with any large-scale language model, including open-source models.'
GitHub - sashiko-dev/sashiko: Agentic review of Linux Kernel code changes · GitHub
https://github.com/sashiko-dev/sashiko?tab=readme-ov-file
Related Posts:
