International Edition
Latest News
Technology

Google’s Sashiko AI Catches 75.8% of Missed Kernel Bugs, Linus Torvalds Endorses Project

Google's Sashiko AI, running the Gemini 4 Carbon model, catches 75.8% of missed Linux kernel bugs on a benchmark of nearly 1,000 real flaws. Google Cloud Blog reported that 241,000 patch reviews have been completed by Sashiko since…

Google's Sashiko AI Catches 75.8% of Missed Kernel Bugs, Linus Torvalds Endorses Project

Google's Sashiko AI, running the Gemini 4 Carbon model, catches 75.8% of missed Linux kernel bugs on a benchmark of nearly 1,000 real flaws.

Google Cloud Blog reported that 241,000 patch reviews have been completed by Sashiko since March 2026, averaging more than 1,200 checks each day.

The Linux kernel spans more than 35 million lines of code and accepts over 85,000 changes every year. Code review remains the main barrier to development speed because maintainers must manually catch memory leaks, deadlocks, and security bugs before changes land in production environments powering billions of Android phones, cloud providers, and supercomputers.

To help maintainers handle the load, the Google AI and Infrastructure team built Sashiko, an open-source agentic code review system powered by Gemini, and donated it to the Linux Foundation earlier this year.

On a benchmark built from nearly 1,000 real Linux kernel bugs that passed human review and later required fixes, Sashiko running the Gemini 4 Carbon model catches 75.8% of the bugs human reviewers missed when operating at a high thinking level, according to Google Cloud Blog.

Linus Torvalds endorsed the project in a mailing-list discussion on lore.kernel.org, writing that the tool increases patch trust. “As you say, Sashiko tends to be quite good. Even when it flags something unnecessarily and there’s no actual problem, there’s likely to be a reason for the question,” Linus Torvalds said on lore.kernel.org.

Sashiko lessons from building an agentic code review system for linux
Photo: wpnews.pro

Reviewing complex codebases requires understanding how a system operates far beyond the specific lines being changed. Sashiko separates finding candidate bugs from verifying them, requiring each review pass to record what it ruled out. The system first gathers surrounding context and domain guides before a planner sets reasoning budgets and assigns specialist passes.

Up to seven specialist sessions run in parallel, focusing on commit intent, implementation match, control flow, resource lifetimes, locking, security, and hardware. Each session averages about 100 tool calls per review using git tools to examine the repository, separating real concerns from investigated candidate issues. A lead verifier then audits both lists together, passing clear-cut bugs through while dropping verified dismissals.

Google's Sashiko AI Catches 75.8% of Missed Kernel Bugs, Linus Torvalds Endorses Project
Photo: thetechstreetnow.com

“Now, I think the other side of AI – and this is where Sashiko comes in – is that it basically does end up being a tool that can then increase the trust in the patch,” Linus Torvalds said on LKML. “So AI giveth, and AI taketh away. But the basic issue shouldn’t be AI per se, it should be that notion of ‘trust’.”

Metric Count
Patch reviews completed 241,000
Mainline commits crediting Sashiko 2,000+
2026 kernel CVEs crediting Sashiko 528
Post-merge fix rate (Linux 7.2) 2.2%

Opted-in kernel subsystem lists now exceed 75 active instances, leading to more than 2,000 mainline commits and 528 Linux kernel CVEs crediting Sashiko in 2026, according to Google Cloud Blog. The 60-day post-merge fix rate dropped from a 3.4% baseline down to 2.2% in Linux 7.2, Google Cloud Blog reported.

Building the system demonstrated that benchmark-driven development outperforms raw intuition, as developers evaluated every major change against a benchmark of nearly 1,000 historical kernel errors, Google Cloud Blog stated. The data also showed that coding agents struggle to review their own generated code, Google Cloud Blog reported.

Engineers are increasingly integrating local Sashiko instances into their coding-agent loops to check patches against community review rules before human maintainers ever see them, according to Google Cloud Blog. Developers can access the project source code on GitHub or through sashiko.dev to run the verification pipeline within their own development workflows, Google Cloud Blog reported.

About the author: Technology Desk

Technology Desk at Archynewsy.