Ai is now reviewing linux kernel code — and outperforming humans
An AI agent called Sashiko is now scanning every patch proposal hitting the Linux Kernel Mailing List, and the numbers coming out of Google's internal tests are hard to dismiss. Built by Roman Gushchin, a Linux kernel engineer at Google, Sashiko caught 53% of bugs from an unfiltered set of 1,000 recent upstream issues — bugs that, according to Gushchin himself, had a 100% pass-through rate on human review. Every single one slipped past human eyes.
What sashiko actually does — and why it matters now
This is not the kind of AI integration that Microsoft has been pushing into Windows 11 — no Copilot button, no generative UI chrome. Sashiko operates at a lower, more consequential layer: it reads kernel patch submissions and flags potential defects before they land in production. It functions as a specialized code review agent, built specifically for the kernel's architecture and idioms, not as a general-purpose assistant awkwardly bolted onto a development workflow.
Gushchin has been running it quietly inside Google for some time. "We've been using it internally at Google for a while and it has helped us discover a large number of real issues," he wrote on LinkedIn. The tool runs on Gemini 2.5 Pro and is now publicly available, covering all LKML submissions — infrastructure and token budget provided by Google.
53% Sounds modest until you read the fine print
Catching just over half of bugs in an unfiltered corpus might not sound like a revolution. But the context reframes it entirely. The benchmark set wasn't curated or cherry-picked — it was raw, recent, real-world upstream issues tagged as fixes. And the human baseline for that same set was zero. Not low. Zero.
That gap is where Sashiko earns its place. It is not positioned as a replacement for human maintainers — Gushchin is explicit about that — but as a layer that catches what fatigued or overloaded reviewers miss. Given that the Linux kernel underpins everything from Android devices to the majority of the world's server infrastructure, "what reviewers miss" is not a trivial category.
Google's fingerprints are all over this — and that's worth noting
The tool is open and publicly accessible, but Google is carrying the operational cost. Gushchin acknowledged the company's "generous commitment to provide the budget and token infrastructure" for the project. That's not a footnote. Running continuous AI review across all of LKML — one of the highest-volume technical mailing lists in open source — is not cheap, and Google is footing the bill.
Adoption is already visible. Maintainers from Qualcomm, Nvidia, and other major contributors are already using Sashiko on their patch queues. For a tool that only recently went public, that uptake is fast.
The Linux kernel has always been a proving ground for what rigorous, distributed, human-driven engineering looks like at scale. Sashiko doesn't disrupt that model — it stress-tests the assumption that human review alone is sufficient. Given a 100% miss rate on the bugs it caught, that assumption was already broken. The tool just made it visible.