Torvalds: AI fixes are driving Linux 7.4, and in BPF the fixes now outpace bug reports

What's going on

Linus Torvalds, the creator and chief maintainer of the Linux kernel, released the second test build of the upcoming Linux 7.4 kernel on Sunday, November 8, 2026. It's called 7.4-rc2, where "rc" stands for release candidate: a pre-release that volunteers are meant to test before the final version ships. These announcements on the kernel mailing list are usually routine. This time, though, one of Torvalds' remarks got people talking. He writes that a large share of the changes come from fixes prompted by AI. And in one part of the kernel, the BPF subsystem, its maintainer says more fixes are now coming in than new bug reports.

That matters because many open-source projects have spent months complaining about a flood of AI-generated bug reports, many of them worthless. If fixes are now keeping up with discoveries, the balance could start to shift, at least for the kernel.

What Torvalds actually wrote

In his post to the Linux Kernel Mailing List, Torvalds sums it up:

And yes, a lot of this has been driven by various AI fixes. I think I'll stop repeating that, and we can just take it as a given. But this time Alexei claims that at least on the BPF side, the fixes are now coming in faster than the new reports, and that sounds great if true.

Alexei Starovoitov is the maintainer, the person responsible for the BPF subsystem. According to MLQ.ai, he alone has 40 commits in the rc2 changelog, mostly BPF fixes plus matching self-tests. Torvalds stays cautious, though:

I think most other areas haven't reached that point yet, but it's at least a sign that we might be starting to see the turn. Don't hold your breath, but maybe keep your fingers crossed that we'll get there in the not too distant future.

BPF stands for Berkeley Packet Filter. It started out as a way to filter network packets. Today, (e)BPF is a kind of mini runtime inside the kernel where small, verified programs can run safely, for things like networking, monitoring or security. Because that code runs deep in the system, it's security-critical and an obvious target for automated bug hunting.

A big rc2 after the biggest rc1 ever

Torvalds calls the release "a big rc2 to match the biggest rc1 ever." The merge window is the roughly two-week period when new features go into the kernel. It closed with rc1 on November 1. According to MLQ.ai, it included "almost 20,000 commits compared to the more usual 15,000 or so." Torvalds wrote at the time that it was "full of fixes for issues found by AI (and increasingly the patches were also suggested by AI, even if they often get rewritten)." He said he'd rather see this AI-driven work land during the merge window than as extra fixes during the stabilization phase, calling that "a much healthier model." Torvalds added that some of the size may also be because developers expect 7.4 to become the next LTS release, meaning a version with long-term support over several years.

If rc2 looks huge at first glance, that's mostly down to automated cleanups done with Coccinelle. Coccinelle is a tool that searches source code for fixed patterns and rewrites them automatically, like a very smart find-and-replace. Torvalds held off on these patches during the merge window because he was traveling, and he openly admitted he didn't feel motivated afterwards either. The scripted changes touch thousands of files, but most are one- or two-line tweaks and, in his words, "really very straightforward."

Where the changes went

Strip the Coccinelle noise out of the diffstat (the summary of which files changed and by how much), and Torvalds says rc2 looks perfectly normal:

OSTechnix also reports early community testing: one tester said rc2 boots without problems on a Dell laptop. As always, release candidates belong on test machines, not production servers. According to Greg Kroah-Hartman, who looks after the stable kernel releases, the final Linux 7.4 release is targeted for December 20, 2026. Until then, rc3 through what will probably be rc7 will arrive weekly.

Background: the AI flood in open source

Tom's Hardware puts the news in a wider context. Many open-source projects have only a handful of volunteers sorting through incoming bug reports, a job known as triage. Since AI tools started churning out reports at scale, projects have been complaining about low-quality "junk" submissions. Some have shut down their bug bounty programs, which pay rewards for finding security holes, or now demand extra proof before they accept a report.

At the same time, AI tools are also finding plenty of real bugs. Tom's Hardware gives examples: one Firefox release in April fixed 423 security bugs, Chrome found and fixed 1,072 bugs in June, and Node.js closed 50 vulnerabilities. MLQ.ai also points out that OpenAI recently said its GPT-6 Cyber model had found 948 exploits, meaning exploitable vulnerabilities, in the Linux kernel. That's a vendor claim the sources don't verify.

The kernel is a special case here. It has far more maintainers and a bigger triage operation than most projects, and the bar for contributions is high. Torvalds has repeatedly argued against rejecting AI-written contributions across the board, as long as they go through the same quality control as human-written code.

Our take

The statement is an interesting signal, but it isn't proof. Torvalds is passing along an observation from Starovoitov, and he hedges it himself ("if true," "might"). The sources give no actual numbers on how many AI reports and how many AI fixes came into the BPF subsystem. It's also unclear what exactly counts as an "AI fix." Torvalds himself says AI suggestions often get rewritten by humans.

Still, the news points to a shift. Until now, AI in open-source projects has mostly been seen as a source of extra work. If the same tools also help with the fixing, the burden could at least partly balance out. For smaller projects without big maintainer teams, that's still a long way off, because the kernel can simply afford to sort through the flood better than they can. And the fact that Torvalds now treats AI involvement as "a given" shows one thing above all: in the world's most important open-source project, AI-assisted development has become everyday reality.