Torvalds AI Code Disclosure: a Cybersecurity Check

October 2026 update: the man who mocked disclosure rules used AI himself

• Torvalds shipped AI-assisted kernel work: on October 11 a branch named noref-filemap-read-with-llm appeared in his tree — an eight-patch series speeding up small page-cache reads (his benchmark: roughly 10M to 30M reads/sec on his laptop). His cover letter said the code, test programs, review, and first draft of the text were produced with Claude Code under his supervision. Not merged as of mid-October; a possible Linux 7.4 merge-window candidate.
• He practices what he preaches: in a separate assoc-array-fixes tag, he used LLM tools to make sense of obscure code and fixed bugs dating back to the library's first commit in 2013 — writing in the tag message that he "did what I preach."
• The deprecation is now in the docs: on October 14, documentation maintainer Jonathan Corbet applied Sasha Levin's patch removing the Assisted-by material (for docs-next). The tag is being deleted; the human's responsibility isn't — the Developer Certificate of Origin and the submitter's own sign-off remain, and maintainers now prefer tool-use detail in the cover letter.
• The spine of the whole story: the disclosure tag died, but accountability didn't. Nobody can hide behind "the AI wrote it" — the human still owns every line they submit.

What happened

• The tag: since January 2026, kernel contributors using "any sort of advanced coding tool" were asked to declare it with an "Assisted-by" tag in their commits.
• The rule that died: the policy lived in Documentation/process/coding-assistants.rst (shipped with kernel 7.0, early 2026). Its ground rules: AI agents must never add a Signed-off-by — only the human submitter certifies the Developer Certificate of Origin; the human reviews all AI-generated code and takes full responsibility; code must be GPL-2.0-only; patches carried an Assisted-by line naming the model and analysis tools.
• The reversal: at the 2026 Maintainers Summit in Prague (October 8, invitation-only, roughly 30 kernel developers), LLMs dominated the day and the room agreed the tag had outlived its purpose (LWN). Kernel developer Sasha Levin (Nvidia) posted a patch on October 9 removing the Assisted-by references; the thread passed 50 messages with replies from Greg Kroah-Hartman, Kees Cook, Jonathan Corbet, and Theodore Ts'o, with new posts still arriving October 16. The Register reported the outcome October 14.
• Why it failed in practice: by the 7.3 cycle only about 10% of non-merge commits carried the tag even though LLM use is far more common — so absence said nothing about how a patch was made. Christian Brauner complained the git history was effectively functioning as free advertising for AI companies; Kees Cook called the tag as redundant as naming one's editor. Red Hat's Jeff Layton backed removal too — not everyone follows the tag, and it's unclear why the kernel should be collecting that information at all. Maintainers now prefer tool-use detail in the cover letter.

What Torvalds said (attributed, not endorsed)

• On the disclosure rules: replying on the kernel mailing list on October 13 to Laurent Pinchart, Torvalds called them "utterly stupid" — understanding code, he argued, has very little to do with the tools used to produce it. He mocked the logic by suggesting contributors drop their compilers and editors and just write in assembler.
• On stage in Prague (Open Source Summit Europe, October 9, with Dirk Hohndel): he quipped that "bots talking to bots" is now real — agentic review systems like Sashiko review patches on the mailing list, and some subsystem maintainers expect submissions to have passed through them first.
• The reversal: Torvalds previously said AI was "just not very good." Now he says he likes it for things he's bad at — like building a modern interface for his guitar-pedal project in C — called vibe coding a wonderful way to find joy in programming, and described AI as a "gateway drug" for new programmers.

The two-sided security picture

• AI finds real bugs: Torvalds says AI tools now surface genuine security flaws, including in drivers untouched for 20 years. On the whole, he believes AI "is improving our code base."
• But it stresses the humans: AI is "also stressing maintainers out to the point that it's a problem" — roughly three-quarters of the Maintainers Summit talks were about making AI tooling less stressful. When bots write the patches and bots review them, the human review layer gets thinner exactly when it matters most.
• His caution: be very careful with AI "when you're doing something real and important."
• The other camp: not everyone agrees with Torvalds — System76 banned AI-generated content from its COSMIC desktop, and NetBSD has banned AI-generated code since 2024.

Cybersecurity FAQ

Honest scope

• This page summarizes reporting on the kernel disclosure debate; it is not legal or compliance advice on AI-use policies.
• Torvalds' positions are his, attributed to his mailing-list posts and stage remarks — the page presents the opposing camp (System76, NetBSD) fairly and endorses neither side.
• The "utterly stupid" phrasing is Torvalds' own, quoted in the headline as reported; all other content is paraphrased.
• Source: Ana Maria Constantin, "Torvalds calls AI disclosure rules 'utterly stupid' as Linux drops its tag", The Next Web, Oct 16, 2026 (summarized, not quoted). Also drawn on (summarized, not quoted): The Register, Oct 14, 2026 (Levin's patch thread and summit consensus); LWN (Maintainers Summit report); Phoronix via secondary reporting (Torvalds' Claude Code branch and benchmarks).