Linux maintainers have dropped the rule that made patch authors disclose AI coding tools, and Linus Torvalds has posted his own LLM-assisted kernel patch. He says a dedicated synthetic benchmark ran 23 times faster, and that no real workload will see a gain of that size.

Tom's Hardware's account of the patch and PC Gamer's account of the tag are the same week's argument, reported from the kernel mailing list and from The Register. At the start of 2026 the project added an Assisted-by tag. The documentation said that if an advanced coding tool helped write a patch, the author had to acknowledge it, so the project could track how AI was showing up in kernel work. Maintainers have now taken that tag out.

Torvalds wants the author to understand the code

Contributor Laurent Pinchart objected to the removal. Torvalds answered on the list that he does not understand a rule of the form "don't use AI for X." His standard, he wrote, is that people do their best and understand what they are doing, and that this understanding has very little to do with which tools they used. He compared a ban to telling contributors they may not use compilers and must write in assembler to prove they know what the machine is doing. "Do you not see how UTTERLY STUPID those kinds of rules would be?" PC Gamer quoted him. The tag, in his view, does not help a maintainer tell a careful patch from a careless one, because both kinds can be written with an LLM in the loop.

Tom's Hardware put the same stance next to a comment he made last July: he was willing to "absolutely put [his] foot down" on people who reject the tools outright. He pointed at an automated review system the kernel already runs, which he said beats human reviewers at catching basic mistakes that should not have been mailed in. He has treated the model as another tool in the set, next to the compiler and the editor.

The 23x figure is a synthetic test in a development branch

The patch is Torvalds applying that rule to his own code. Tom's Hardware said he wrote a draft of the underlying fix in early 2024 and has run that version on his own laptop since. He then tried to solve the same problem with an LLM, posted the result on a development branch, and reported the 23-fold gain on a benchmark built for the test. He added that the number will not show up under a real load. The article does not name the subsystem in enough detail to treat the speedup as a general kernel result, and it should not be read as one.

Anthropic tested a worm that copies model weights onto cheap hardware. That was a lab measuring what a model can do when it gets out. The kernel story is the other direction: maintainers deciding what to do with code a model helped write. OpenAI has scrapped a model release over safety concerns, which is a company pulling a model. Torvalds is not pulling the tool. He is refusing a label on the patch.

The security list is the part that got worse

Tom's Hardware also reported the cost on the other side of the same tools. Earlier this year the Linux security mailing list became unmanageable because agents are good at spotting common flaw patterns and filing them. The project has been fixing on the order of 2,000 CVEs per release. Old drivers, with a small installed base and a large pile of reports, are being dropped so the maintainers can keep up. That is a separate decision from the Assisted-by tag, and it is the one Torvalds's pro-tool comments do not erase. The tag tracked authorship. The CVE pile is the review queue.