Google's AI catches 53% of Linux bugs humans miss
Google engineers built Sashiko, an AI that reviews Linux kernel code and catches 53.6% of bugs that every human reviewer missed. It's free and open source.
A team of Google engineers just launched an AI system that does something no human code reviewer has managed: it catches more than half the bugs slipping into the Linux kernel — the software that powers everything from Android phones to cloud servers to the International Space Station.
The tool is called Sashiko (named after Japanese reinforcement stitching), and here's the jaw-dropping number: when tested against the last 1,000 real bugs found in the Linux kernel, Sashiko caught 53.6% of them — and 100% of those bugs had been missed by every human reviewer.
How an AI outperforms thousands of kernel developers
The Linux kernel gets contributions from thousands of developers worldwide. Every code change goes through a rigorous human review process on public mailing lists. Yet bugs still get through — some causing crashes, security holes, or data corruption.
Sashiko monitors those same mailing lists automatically and runs every proposed change through a 9-stage AI review process:
What Sashiko checks — from big picture to tiny details:
1. Does this change break the overall design?
2. Is the code doing what it claims?
3. Are there logic errors in how the code runs?
4. Could this leak memory or crash the system?
5. Are multiple processes going to fight over the same data?
6. Security audit — buffer overflows, race conditions, privilege issues
7. Hardware-specific problems (register access, DMA mapping)
8. Double-checking its own findings and rating severity
9. Writing a clear report with recommendations
The false positive rate (times it flags something that isn't actually a bug) is under 20% — meaning 4 out of 5 warnings are real problems worth investigating.
Who built it, and who's paying for it
The project was created by Roman Gushchin, a Google engineer on the Linux kernel team. Google is funding both the computing power and the AI token costs (the fees charged every time the AI processes code).
The Linux Foundation now officially hosts the project, and it's fully open source under the Apache 2.0 license. Anyone can inspect, modify, or build on it.
Under the hood, Sashiko runs on Google Gemini 3.1 Pro by default, but it also works with Anthropic's Claude — including support for Claude's 1-million-token context window (meaning it can analyze enormous amounts of code at once).
Why this matters beyond Linux
If AI can catch bugs in the Linux kernel — one of the most carefully reviewed codebases on Earth — imagine what it can do for the code at your company that gets far less scrutiny.
For developers: Sashiko is open source and you can run it on your own projects. It requires Rust 1.86+ and an API key from Google or Anthropic.
For engineering managers: This is proof that AI code review isn't a gimmick. A 53.6% bug catch rate on problems that slipped past expert human reviewers is a genuine safety net.
For everyone else: The software running your phone, your car's infotainment system, and most of the internet's servers just got a new quality checker. Fewer kernel bugs means fewer crashes, fewer security vulnerabilities, and more reliable tech.
Try it yourself
You can browse Sashiko's live reviews at sashiko.dev — it's reviewing kernel patches in real time right now.
To run it locally:
git clone --recursive https://github.com/sashiko-dev/sashiko.git
cd sashiko
# Configure Settings.toml with your LLM API key
cargo build --release
cargo run --releaseThe project is still young (81 GitHub stars), but with Google's backing and Linux Foundation governance, it's worth watching closely.
The big picture: Human code review isn't going away. But AI just proved it can be the safety net that catches what even the best developers miss. Sashiko's creator put it best — the tool is designed to augment human reviewers, not replace them. Given that it found bugs in the most scrutinized codebase in the world, that's a pretty compelling argument.
Related Content — Get Started with Easy Claude Code | Free Learning Guides | More AI News
Stay updated on AI news
Simple explanations of the latest AI developments