On August 3, 2026, Greg Kroah-Hartman, the maintainer of stable Linux branches and the staging subsystem, announced new rules for accepting changes into the drivers/staging directory. This was prompted by a recent influx of LLM-generated patches. Now, fixes prepared by language models for this part of the kernel will be automatically rejected, except for confirmed fixes for real vulnerabilities.
The drivers/staging directory is used to host drivers and other components that do not yet meet the requirements of the mainline Linux. According to the official kernel documentation, such code stays in staging until it requires additional work; once it is refined, it can be moved into the main part of the kernel, while stagnant components are removed over time.
However, Kroah-Hartman pointed out another task of this subsystem: drivers/staging primarily exists as a place where new developers can learn to participate in kernel development. Here, many relatively simple tasks—code formatting fixes, transitions to new APIs, and minor redesigns—are intentionally left. By completing these tasks independently, newcomers learn how to prepare patches, communicate in mailing lists, and undergo the review process.
In the maintainer's opinion, the automatic rejection of such code contradicts the very purpose of staging. He noted that developers could fix all formal shortcomings using automated tools if that were the main goal. Instead, some simple tasks are preserved so that people can learn from them. Thus, Kroah-Hartman intends to automatically reject any patches created by LLM for the drivers/staging subsystem.
Attempts to avoid indicating the use of a language model will not help either. Kroah-Hartman stated that such changes are usually easy to recognize and warned authors that deliberately concealing the origin of a patch could be seen as an attempt to deceive the maintainers. The goal is for people to learn, not to try to outsmart the maintainer, the developer explained.
The only exception is made for fixing actual security vulnerabilities. Such a patch is allowed to be submitted, but the author is required to test it on the actual hardware for which the driver is intended and to detail the testing conducted. Additionally, the submitter must be prepared to demonstrate that the issue is indeed reproducible and may affect users. According to Kroah-Hartman, "at least one third" of the issues identified by modern LLMs turn out to be erroneous or lead to harmful changes..
Kroah-Hartman compared the drivers/staging to "a gym where one can learn and hone skills."Language models, he says, are gradually becoming a useful tool for "heavy lifting," but they should be used by specialists who already possess sufficient knowledge to evaluate the results. Thus, the new rule does not impose an outright ban on the use of AI in Linux development: it only pertains to the drivers/staging tree supervised by Kroah-Hartman and primarily concerns automatically generated cosmetic fixes.
Source: linux.org.ru
