Linux Kernel AI Code Review Makes Huge Release Candidates Normal

Torvalds stops fighting the flood

The Linux kernel releases test builds every week, and their size used to tell maintainers how healthy a cycle was. That measure has stopped working. When Linus Torvalds released Linux 7.2-rc7 on August 9, 2026, he said he was not thrilled about how big it was. He accepted it anyway as the new normal: a large number of fixes, many of them found by various AI review tools.3 The release candidate carried more than 400 fixes from over 230 contributors.5 TechRadar noted that this is the kind of volume usually seen early in a cycle, not in the last candidate before a stable release.8

Two months later, the phrase has become routine. Torvalds released Linux 7.3-rc6 on October 4 and said its commit counts were normal for the "new AI normal," with many small fixes to error-handling paths.23 Drivers again made up much of the patch, across GPU, networking, USB, TTY, IIO and sound code.25 Torvalds is no longer complaining about the volume. He now uses it as the baseline.

The main point is this: AI review tools have permanently raised the amount of fix traffic in the kernel. The project's job has changed from keeping that traffic out to running a process that can absorb it.

From irritation to acceptance

The path to this point was rough. In May, with Linux 7.1-rc4, Torvalds called the kernel's private security mailing list "almost entirely unmanageable." He blamed duplicate reports from researchers running the same AI tools against the same code.4 His reasoning was simple: if an AI tool can find a bug, other people's tools will find it too, so keeping those reports private makes little sense.1 New kernel documentation now says AI-found bugs should be treated as public and sent straight to the relevant maintainers, with a verified way to reproduce them.4

A week later, the 7.1-rc5 release was bigger than any recent rc5. Torvalds blamed a wave of trivial fixes and said several patch series had been triggered by AI code review.2 He promised to become "more hardnosed" and to turn away late-cycle pull requests that did not fix regressions (bugs that break something that used to work).2 Coverage at the time read this as frustration with AI reviewers making a trivial-fix problem worse.9

By August, the tone had changed. Torvalds said nothing in 7.2-rc7 looked scary and saw no reason to delay the final release.5 In September, when 7.3-rc2 came in unusually large, he joked that everyone would "blame it on AI" whether or not that was the real cause.27 Outlets disagree on how seriously to take that remark. Betanews and gHacks note that Torvalds never confirmed AI was behind the rc2 growth.2629 Other coverage treats it as more evidence of the AI-driven trend.30 The careful reading is that rc2 was a joke and not data, while the rc7 and rc6 remarks were direct statements of how things now work.

Review, not authorship — mostly

The biggest difference between outlets is about what the AI is actually doing. TechRadar says Torvalds credits AI review, not AI authorship. In its account, review agents point at existing code and produce bug reports, and humans write and submit the fixes.8 Betanews gives a similar picture: the tools review code and file reports, people triage every finding, and policy requires a human to sign off on any AI-assisted contribution.5

That picture is accurate for review, but it leaves out how much code AI now helps write. Phoronix reported in May that GitHub Copilot and Claude Code were the most frequent AI contributors to kernel development. It pointed to patches generated or co-written by those agents in the Intel Xe driver, AMD display code, SMB, Netfilter, io_uring, Bluetooth and other areas.18 One analysis says roughly 16% of patches in recent cycles were tagged as AI-assisted. That number should be treated as approximate, since the outlet itself describes it as reported, not measured.21

The 7.2-rc7 cycle includes one fix that shows both roles at work. Syzbot, Google's automated kernel fuzzer (a tool that feeds random inputs to code to find crashes), flagged an eight-year-old race condition in ptdump that could lead to a use-after-free bug. Developer David Carlier wrote a first fix with Claude Opus 4.8 helping trace the root cause, and memory-management maintainer Lorenzo Stoakes reworked it before it was merged.5 Automated tools found the bug, an AI model helped diagnose it, and humans wrote and approved the final fix. That pattern describes the "new normal" better than either "AI review" or "AI authorship" alone.

Claude Code's growing kernel footprint

Claude Code shows up throughout the kernel's AI story. In April, Anthropic researcher Nicholas Carlini described using it to find a remotely exploitable heap buffer overflow in the kernel's NFS driver that had been there since 2003. He has five confirmed kernel vulnerabilities in total, plus several hundred crashes he has not yet validated.17 His method was close to brute force: a script that went through every source file and asked Claude Code to look for vulnerabilities as if it were in a capture-the-flag contest.17 He also said earlier models found only a small share of what Opus 4.6 found. That fits the sudden jump in report quality maintainers describe.17

One contributor's account shows the routine side of this. The developer used Claude to review legacy Intel Wi-Fi driver code and draft patches and tests, then had Codex challenge the claims. That cross-check shrank the patch series and forced one patch to be rewritten.12 The fixes were tested with KUnit and KASAN under QEMU, disclosed with an Assisted-by credit, and accepted upstream.12

The agent tools are now changing kernel process too. NVIDIA engineer Sasha Levin proposed an AGENTS.md file, the standard instruction file many coding agents load automatically, so agents would find the kernel's AI rules. In one test without the file, an agent added a Signed-off-by tag by itself, which only humans are allowed to do.16 Theodore Ts'o objected that it would waste tokens, and Laurent Pinchart rejected AI-assisted development entirely. No consensus had been reached.16

The rulebook and the Sashiko fight

The acceptance rests on a formal policy. The kernel's AI Coding Assistants document shipped with Linux 7.0.15 It bars AI agents from adding Signed-off-by tags. It requires an Assisted-by tag naming the model and tools used, and it makes the human submitter fully responsible for the code.11 Torvalds has also warned that people who submit low-quality AI code will not label their patches honestly. That means disguised, plausible-looking patches are the real risk, not obvious junk.11

AI code review came to a head in July over Sashiko, an open-source tool that uses large language models to review patches. Critics said its false positives made it a burden.10 Torvalds answered that Linux is "not one of those anti-AI projects" and that objectors could fork the kernel or walk away.30 Sashiko reports finding bugs in 53.6% of a sample of 1,000 recent commits that later needed fixes, all of which had passed human review. It used Gemini 3.1 Pro and estimates its false-positive rate at around 20%.6 Those are the tool's own numbers, and the false-positive estimate comes from limited manual checking.6 Whether Sashiko stays will likely depend on whether maintainers find it saves more time than it costs.10

The bill comes due downstream

The costs are real. Greg Kroah-Hartman reportedly showed at Kernel Recipes 2026 that the kernel fixed about 500 CVEs (publicly tracked security vulnerabilities) per release from 6.9 to 6.19. That rose to about 1,000 from 7.0 onward, more than 1,500 in 7.2, and possibly over 2,000 in 7.3.30 At the start of the 7.3 cycle he warned it would be "a rough -rc cycle." He said that even clearly wrong AI patches take time to reject, because he does not want to throw out real fixes by mistake.26 Some networking fixes in 7.3 were delayed by the volume of AI-driven reports.28

The effects are spreading to Linux distributions. Canonical moved Ubuntu to a single two-week update cycle and said the explosion in CVEs was driven by AI.22 Analysts also link AI reports against barely used hardware to old drivers being removed faster.21 The underlying problem is that only a few hundred core maintainers review the bulk of patches.21

What to watch

Linux 7.3 is due around October 18, with October 25 as the fallback if an rc8 is needed.2625 The maintainers summit and Linux Plumbers Conference take place during the final week.23 They are the natural places to settle open questions like AGENTS.md and how much triage maintainers can handle. Torvalds has already accepted AI review. The open question is whether the kernel's human review capacity can keep up with it.