By ByteBot
Linux 7.2-rc7 landed on August 9 carrying 412 file changes from over 230 contributors — an anomaly for a late release candidate. Linus Torvalds acknowledged it plainly: the volume is “the new normal,” driven by “review by various AI tools.” He’s not exactly thrilled. But he isn’t pushing back either. The creator of the world’s most critical open-source project has accepted AI as a permanent fixture in kernel development.
The Numbers Don’t Lie
A typical rc7 carries 150 to 200 file changes. This one had 412. That’s not noise — it’s a structural shift. RC6 was already described as the largest rc6 in kernel history. Nearly 29% of recent commits carry “Fixes:” tags, close to double the historical baseline of around 15%. AI tools scanning the codebase continuously are surfacing bugs faster than the release cycle was designed to absorb.
Torvalds put it this way: “I can’t say that I’m exactly thrilled about the size of this all.” He’s right to be ambivalent. The fixes are legitimate. Nothing looks scary enough to delay the August 16 stable release. But the pace of incoming patches has changed, and it isn’t going back.
Meet Sashiko: The Reviewer That Never Sleeps
The tool sitting at the center of this is Sashiko, an agentic code reviewer built by Google’s Roman Gushchin and launched in March 2026. Named after a Japanese embroidery technique — decorative reinforcement stitching — Sashiko reviews every patch submitted to LKML and several other kernel mailing lists using Gemini 3.1 Pro. Greg Kroah-Hartman, the stable kernel maintainer, endorsed it publicly. It’s open source at github.com/sashiko-dev/sashiko.
Its performance is hard to argue with. Tested against 1,000 recent upstream issues tagged with “Fixes:”, Sashiko found approximately 53% of the bugs. Human reviewers had missed all 1,000. That asymmetry is worth sitting with: a tool running on a cloud model caught bugs in code that hundreds of experienced kernel contributors walked past.
What AI Is Actually Finding
RC7 itself included a fix for an 8-year-old memory management race condition — a use-after-free that had been quietly lurking in production kernels. For context: in April 2026, Claude Code found a 23-year-old heap overflow in the NFSv4.0 LOCK replay cache. A 112-byte buffer was receiving 1,056 bytes through an owner ID field. Remotely exploitable. Sitting there for two decades. AI found it in hours.
According to Phoronix’s RC7 analysis, the kernel security list now receives 5 to 10 valid AI-generated security reports daily. That’s a significant signal increase. The reports aren’t slop anymore — maintainers confirm that quality has improved substantially since mid-2025.
The Governance Layer: Assisted-by
The Linux kernel community didn’t just accept AI tools and move on. In April 2026, after months of debate, it codified the rules in a document called coding-assistants.rst. The core requirement: if AI helped write or review a patch, declare it with an Assisted-by tag.
Assisted-by: Claude:claude-opus-4-6 coccinelle sparseAI systems cannot use Signed-off-by, which carries legal weight in the kernel’s development process. The human who submits a patch bears full responsibility — for bugs, security issues, and licensing. As Tom’s Hardware reported when the policy landed, this is the kernel’s way of saying yes to AI tools without surrendering human accountability.
The Unexpected Casualty: Old Code Is Getting Deleted
Here’s the part that doesn’t get discussed enough: AI-driven scrutiny is forcing legacy code out of the kernel faster than any cleanup effort in recent memory. In April, Torvalds merged a cleanup that stripped 138,000 lines of ISDN and PCMCIA-era code. In August, Greg Kroah-Hartman deleted the Moxa Intellio serial driver — 2,148 lines written in 1999, hardware discontinued in 2017 — because LLMs were “poking at the code and find interesting things that would waste maintainers’ time.”
That logic is worth understanding. The driver wasn’t actively broken. But maintaining it under constant AI scrutiny costs more than retiring it. That pattern will repeat across the kernel and, eventually, across every major open-source project that AI tools start reviewing at scale.
What This Means for Developers
Torvalds made his position clear in July: “Linux is not one of those anti-AI projects.” He told critics to fork the kernel if they disagreed. That’s not enthusiasm — it’s pragmatism. AI tools are catching real bugs at a rate human review can’t match. The cost is more triage work for maintainers, not less overall effort.
For developers contributing to the kernel, the calculus is straightforward. Use the Assisted-by tag when AI helps. Expect AI-flagged issues in code you haven’t touched in years. If you’re maintaining a legacy driver for hardware discontinued in 2003, start thinking about what happens when Sashiko finds something your morning review missed — and whether fixing it is worth the effort or it’s time to let the driver go. The AI review era in open-source development isn’t approaching. It’s already here, and Torvalds just confirmed it.
