Google’s Open-Source AI Reviewer Sashiko Finds Linux Bugs Humans Miss
November 18, 2026
Google AI & Infrastructure built Sashiko, an open-source, Gemini-powered code-review system, to help Linux kernel maintainers handle a growing patch workload and catch bugs before changes are merged; the project was donated to the Linux Foundation.
Sashiko gathers surrounding code and subsystem guidance, then runs up to seven parallel specialist reviews targeting different bug categories; a lead verifier triages findings, while separate agents investigate difficult cases and require concrete code evidence to accept either a bug claim or its dismissal.
A benchmark of nearly 1,000 real kernel bugs—issues that passed expert review, reached mainline and later required fixes—helped guide the system’s development; using Gemini 4 Carbon at a high thinking level, Sashiko reportedly detected 75.8% of benchmark bugs missed by human reviewers.
After seven months in production, Sashiko had completed about 241,000 reviews, or roughly 1,200 a day, with more than 1,500 kernel developers engaging with its reports; its findings were cited in over 2,000 mainline commits and 528 Linux kernel CVEs in 2026.
The service reviews public Linux mailing-list patches for free and sends reports only to opted-in lists; it also publishes its prompts, tool calls, reasoning steps and ruled-out concerns, with more than 100 contributors involved and over 70 kernel mailing lists opted in.
Among 21,500 tracked mainline commits, patches with clean Sashiko reviews required a follow-up fix within 60 days 0.9% of the time, compared with 7.6% for patches merged despite a critical warning—evidence the article presents as reassurance from clean reviews and a reason to scrutinize critical warnings.
The share of merged commits receiving a follow-up fix within 60 days declined from a 3.4% baseline across Linux 6.2–6.19 to 2.8% in Linux 7.1 and 2.2% in Linux 7.2, a 35% drop from baseline; the authors say the trend is promising but cannot be attributed solely to Sashiko.
The authors’ broader lessons are to favor benchmark-driven evaluation over intuition, have coding agents review work independently rather than their own output, and build trust through transparency and community ownership.
The authors say Sashiko could also be adapted for foundational projects such as GCC, LLVM and Git, since its core review engine is not limited to C or kernel-specific rules.
Summary based on 1 source
Get a daily email with more AI stories
Source

Google Cloud • Nov 19, 2026
Lessons from building Sashiko, an agentic code review system