Google’s open-source AI reviewer Sashiko has become part of Linux kernel development, catching bugs human maintainers missed across 241,000 patch reviews since March. Google’s AI and Infrastructure team built the system, powered by Gemini, and donated it to the Linux Foundation earlier this year.
How It Reviews
Sashiko splits finding bugs from proving them. Up to seven specialist sessions run in parallel, each hunting one class of problem — control flow, resource lifetimes, locking, security, hardware — and each must record what it ruled out, not just what it found.
A lead verifier then audits both lists together, and every contested case gets its own deep-check agent. Neither a bug claim nor a dismissal counts without concrete code proof across all call sites, which catches specialists talking themselves out of real bugs.
What the Data Shows
Sashiko now completes about 1,200 reviews a day across more than 100 kernel mailing lists, with over 75 lists opted in. More than 1,500 developers have replied to its reports on the lists, and their pushback on false alarms has directly improved its prompts and checks.
The system is cited in more than 2,000 mainline kernel commits and 528 of this year’s Linux kernel CVEs, spanning drivers, networking, GPUs and memory management. On a benchmark of nearly 1,000 real bugs humans missed, Sashiko with Gemini 4 Carbon caught 75.8 percent of them.
Our Read
The numbers that matter are the ones about trust. Patches that passed Sashiko cleanly needed a follow-up fix within 60 days just 0.9 percent of the time, against 7.6 percent for patches merged despite a critical warning — an eightfold gap that maintainers can actually use.
Google is careful not to claim the improvement: the 60-day fix rate fell from 3.4 percent to 2.2 percent in Linux 7.2, a 35 percent drop, but other factors shape release-to-release numbers. Agents are finding flaws on their own now, and Sashiko shows the same capability aimed at prevention.
Linus Torvalds praised the reports in a mailing-list discussion, calling some of them deep, subtle and completely correct. His verdict cuts through the ‘AI slop’ debate: a reviewer that earns maintainers’ trust is worth more than a generator that floods them with patches.
Why It Matters
Linux runs billions of Android phones, most cloud providers and nearly every supercomputer, so a measurable drop in bugs landing in mainline releases is a real-world result, not a benchmark score. The kernel is the largest open-source project, and it now runs an AI reviewer in production.
The design also travels. Sashiko’s core engine does not depend on C or on kernel-specific rules, so Google says the same playbook — open verification, community-owned guides, a benchmark mined from real fixes — could apply to GCC, LLVM, Git and other foundational projects.
Background
Coding agents are not great at reviewing their own work, Google argues, because producing a working solution is only a small share of the job in a codebase. Trust in coding tools has been tested this month, after a poisoned extension reached thousands of developers through a leaked vendor key.
Source: Google Cloud Blog


0 Comments