The Linux Staging Subsystem Will Reject LLM Patches — Except for Genuine Security Fixes

Greg Kroah-Hartman, the second-in-command in the Linux kernel development hierarchy, announced a new policy on tools based on large language models on Monday. The catalyst, in his words, was a “flood” of LLM-generated patches for the drivers/staging/ subsystem. From now on such patches will be rejected automatically — with a single exception, which is discussed below.

Why staging exists

Kroah-Hartman points out that drivers/staging/ is meant first and foremost as a training ground for newcomers. It is home to plenty of “low-hanging fruit” — minor code cleanups and API fixes on which a beginning developer can learn the process without pressure or risk: the assumption is that no one seriously relies on the code in these directories (which is why the kernel sets the TAINT_CRAP flag when it is loaded).

This is precisely why maintainers already refuse to accept bulk changes along the lines of “fix all the style issues in this file” — otherwise the point of the subtree would be lost. If the team cared about the quality of this code, it would have been cleaned up in a single day; instead the shortcomings are tolerated deliberately, so that people have something to learn on.

Hence the conclusion: using an LLM to “tidy up” staging directly contradicts its purpose. Kroah-Hartman also has a warning for anyone who tries to hide the fact that a model was used: in his words, machine-generated patches are “very obvious,” and any attempt to deceive a maintainer will be treated accordingly.

The exception for security

There is one exception after all. Modern models, Kroah-Hartman concedes, are fairly good at spotting suspicious places in kernel code. But even the best tools of today and tomorrow produce results that are simply wrong or harmful in at least a third of cases.

For that reason a vulnerability fix found with the help of an LLM can be accepted — but on a strict condition: the author is required to test the patch in advance on real hardware for the driver in question and to describe exactly how they did so. Such testing should filter out false positives, and the author must be prepared to defend the change and prove that it fixes a real bug a user could actually run into.

Kroah-Hartman phrases his metaphor this way: staging is a “gym” for building up skills, and an LLM is fit for “heavy lifting” only in the hands of someone who has already trained and understands what can be lifted and how. In support of this he cites an essay by Bruce Schneier.

Beyond staging, the kernel's position remains unchanged: AI-assisted work is still allowed. Linus Torvalds treats such tools as ordinary tools and emphasizes that the Linux project is “not anti-AI.”

Source: Phoronix