What did Linus Torvalds say about AI fixes in Linux 7.4-rc2?
Linux 7.4-rc2 arrived on Sunday 8 November 2026, one week after what Torvalds called the largest rc1 in the kernel's history. The second test build of the next kernel is normally a quiet checkpoint. This one was not. Linus Torvalds used his weekly release note to describe a large update and to make a remark that has since travelled across the open-source world.
Torvalds called it "a big -rc2 to go with the largest ever -rc1", according to LWN's summary of the mailing-list post. Part of that size was mechanical. He had held back automated Coccinelle clean-up patches during the merge window, then merged them at the start of the rc2 cycle. Those changes touch thousands of files but mostly consist of one-line or two-line edits.
Once that noise is removed, Torvalds said the shape of the update looks ordinary. Roughly a third is drivers, with graphics, networking and sound accounting for most of it. Another third is networking and BPF. The final third covers filesystems, memory management and architecture updates. Then came the sentence that defined the cycle.
"And yes, much of this was driven by various AI fixes," Torvalds wrote. He added that he would stop repeating the point and that readers could take it for granted. The line is short, but it marks a change in how the kernel's top maintainer describes machine-generated work. In this account, AI is part of the fix stream.
He then relayed a claim from Alexei Starovoitov, the co-creator and lead developer of the extended Berkeley Packet Filter subsystem. On the BPF side, Torvalds wrote, fixes are now arriving faster than new reports. He called that "lovely if true", a phrase that keeps the claim at arm's length while acknowledging its significance.
The caution matters. Torvalds did not say the whole kernel has turned the corner. He said most other areas have not reached that point. Even so, he saw it as a sign that the tide may be starting to turn. He also told readers not to hold their breath, but to cross their fingers that other subsystems get there in the not entirely distant future.
Why is BPF the first Linux subsystem to out-fix AI reports?
BPF, often called eBPF, lets sandboxed programs run inside the Linux kernel without changing kernel source. Developers use it for networking, observability, security filtering and performance tracing. It has become one of the most active parts of the kernel, which makes it a natural place for AI bug-finding tools to concentrate their effort.
The subsystem also has a large, engaged maintainer group. Starovoitov is its lead developer, and BPF patches pass through a demanding review process with extensive testing. A narrow, well-instrumented target is easier to fix quickly than a sprawling area such as memory management or filesystem code. That may explain why BPF reached the fix-outpaces-reports point before its neighbours.
The milestone is about rates, not absolutes. If a team fixes ten bugs while AI tools report eight, the fix rate is ahead. The subsystem can still contain unknown flaws, and a later surge of reports could reverse the position. Torvalds's wording reflects that uncertainty. He repeated Starovoitov's claim rather than presenting it as an audited measurement.
BPF's prominence in 7.4-rc2 is visible in the diff. Torvalds grouped networking and BPF together as roughly one-third of the release candidate once automated clean-ups were excluded. That is a large share of a kernel cycle. It also means any change in BPF's bug economics has an outsized effect on the overall patch mix.
There is a practical lesson here for other projects. AI bug hunters tend to produce volume: duplicate reports, low-severity findings and patches that need triage. A subsystem with clear ownership, strong tests and maintainers who can say no can absorb that volume. BPF appears to have reached the point where the incoming work is being closed as fast as it arrives.
Whether that state lasts is the open question. Torvalds's note covers one week of one cycle. A single subsystem's experience is an anecdote, not a trend. But it is an early concrete signal from the kernel that AI-assisted bug discovery might eventually reduce, rather than only add to, maintainer workloads.
How large is the AI-driven patch flood in the Linux kernel?
The scale of recent kernel work is unusual even by Linux standards. LWN reported that the 7.4 merge window brought in 18,591 non-merge changesets. That set a record and came close to matching the patch count of the entire 7.3 development cycle, which had 18,765. The 7.4 cycle was already exceptional before the stabilization phase began.
Stable branches show the same pressure. On 10 November 2026, Greg Kroah-Hartman released Linux 7.3.4 with more than 1,250 patches, 6.18.63 with more than 860, and 6.12.117 with more than 1,530, according to LWN. Those are maintenance updates, not feature releases. Each one still has to be reviewed, tested and backported by people who often work on the kernel in addition to other jobs.
Part of the surge traces to better bug-finding tools. Anthropic published research on Claude Mythos Preview on 7 April 2026. The company said the model could, when directed by a user, identify and exploit zero-day vulnerabilities in every major operating system and every major web browser. It also said it had found thousands of additional high- and critical-severity vulnerabilities that were being disclosed to maintainers.
The Register draws a direct line from that capability to the kernel's patch load. It notes that developers have been able to find more software flaws since Mythos arrived in April. The same report cites a record 974 flaws addressed by Microsoft in September and 1,449 flaws addressed by Oracle in a single month. The kernel, with more than 40 million lines of code, is an obvious target for the same techniques.
AI does not only create work for kernel developers. It also creates work for the people who review their patches. On 10 November, Phoronix reported that Lorenzo Stoakes was stepping away from reviewer roles for the memory management, core, misc and Rust areas. The reason was the noise from the flood of kernel patches produced with AI and large language models.
That human cost is easy to miss in a story about faster fixes. Every AI-generated report still needs a human to decide whether it is real, whether it is a duplicate, and whether the proposed patch is correct. Volume that arrives faster than reviewers can process it becomes a queue, then a backlog, then a reason for experienced people to step back.
Does one subsystem mean the AI bug flood has peaked?
No. Torvalds was explicit that BPF is an exception. "I think most other areas haven't reached that point yet," he wrote, according to The Register. The tide may be turning, but it has not turned everywhere. Memory management, filesystems, graphics drivers and architecture code all have different maintainers, test coverage and report volumes.
A faster fix rate can also coexist with a larger backlog. If AI tools report bugs at three times the old rate and maintainers fix them at twice the old rate, the queue still grows. BPF's milestone says something about the balance in one subsystem. It does not prove that the total amount of outstanding kernel work is falling.
The quality of AI reports remains a variable. Maintainers have complained about duplicate findings and reports that waste review time. Anthropic says it uses human validators before disclosing Mythos findings, and that its contractors agreed exactly with the model's severity assessment in 89% of 198 manually reviewed reports. Even a high agreement rate leaves a meaningful number of reports that need correction.
There is also an incentive problem. AI bug-finding tools are improving quickly, and defenders are not the only users. Anthropic warns that the short-term advantage could go to attackers if frontier labs are not careful about how they release such models. A kernel that fixes AI-found bugs faster is more secure, but it is also responding to a moving baseline.
The most defensible reading of 7.4-rc2 is therefore narrow. One subsystem, BPF, has reached a healthier balance between incoming reports and outgoing fixes. The rest of the kernel has not. Torvalds's optimism is conditional, and he says so in the same message that contains the good news.
What does Linux 7.4 mean for users and distributions?
Linux 7.4-rc2 is a test kernel. It is not intended for production servers, work laptops or any system whose downtime would matter. Release candidates exist so that testers can find regressions before the final release. Torvalds publishes them weekly, and each one includes fixes that arrived after the previous build.
The cycle began with 7.4-rc1 on 1 November 2026, after a merge window that OSTechNix measured at 13.92 days and 19,780 total commits. The rc2 build followed on 8 November. OSTechNix reports that Greg Kroah-Hartman projected 20 December 2026 as the final release date in a 2 November mailing-list reply. That date is a projection and can move if more release candidates are needed.
The release also includes ordinary feature work alongside the AI-driven fixes. Phoronix reports FAT32 optimisation and an AMDGPU HDMI 2.2 improvement in 7.4-rc2. Those changes matter to desktop and embedded users, but they are not the reason this cycle attracted attention. The AI angle is.
For distribution maintainers, the practical effect is more backport work. The 7.3.4 update alone carried more than 1,250 patches. Distributions must decide which fixes to take, test them against their own configurations and ship them to users. A faster upstream fix rate can shorten the time to a security update, but it can also increase the volume that distribution teams must process.
Users who want the newest hardware support usually wait for the final release or for their distribution to adopt it. Users who value stability should look at long-term support branches instead. OSTechNix reports that Kroah-Hartman plans to make Linux 7.3 the next LTS kernel, telling developers to count on 7.3.y unless something unusual happens. That makes 7.3 the branch to watch for long-lived deployments.
The safest way to help is to test release candidates on spare hardware and report regressions with precise details. Kernel testing is not glamorous, but it is the mechanism that turns a large release candidate into a reliable final release. AI can generate more fixes; it cannot replace that feedback loop.
Are AI-assisted patches allowed in the Linux kernel?
Yes, within rules. The kernel's official AI Coding Assistants document says AI tools should follow the normal development process, including coding style and patch submission guidelines. It also sets a clear legal boundary. AI agents must not add Signed-off-by tags, because only humans can certify the Developer Certificate of Origin.
The human submitter carries the responsibility. According to the document, that person must review all AI-generated code, ensure licensing compliance, add their own Signed-off-by tag and take full responsibility for the contribution. The rules do not ban AI assistance. They place the accountability on the person who sends the patch.
That approach explains why Torvalds can describe AI fixes without treating them as a separate category of code. A patch is judged by what it does and by who signs it, not by whether a model helped produce it. The kernel has debated how much disclosure to require, but the underlying principle has stayed consistent: a human is answerable for every change.
The policy also has limits. A review requirement is only as strong as the reviewer's time and attention. If AI tools increase the number of patches faster than maintainers can inspect them, the rule that humans remain responsible becomes harder to honour in practice. That is the tension behind the Stoakes departure and the broader debate about low-quality AI-generated reports in open source.
Other projects are wrestling with the same questions. LWN reported on 9 November that the Git project is considering a vote to settle its own LLM contribution policy. The Linux kernel's experience is therefore a preview, not an isolated case. Every large open-source project will need a way to accept useful AI help without drowning its reviewers.
For contributors, the practical advice is simple. Use AI tools if they help, test the result, disclose what the project requires and sign off only on work you understand. A model can find a bug or draft a fix. It cannot take responsibility for the patch, and the kernel's rules do not pretend otherwise.
What should you watch next?
The next data points will come weekly. Linux 7.4-rc3, rc4 and later builds will show whether the BPF fix rate holds, and whether other subsystems begin to report the same balance. Torvalds's messages are the primary record. Starovoitov's own mailing-list updates will matter too, because the claim currently reaches the public through Torvalds's summary.
Watch the stable branches as closely as the release candidates. The 7.3.y series is the planned long-term support line, according to OSTechNix's account of Kroah-Hartman's plans. Its patch volume will show how much AI-driven fixing is reaching users who never run a release candidate. A drop in stable patch counts would be a stronger signal than a single good week in BPF.
Watch maintainer staffing as well. The Stoakes story is a reminder that open-source capacity is human. If experienced reviewers step back, the kernel's ability to absorb AI-generated work falls. Projects that invest in triage, testing and clear ownership will cope better than those that rely on a few overloaded volunteers.
Finally, watch the tools. Anthropic's Mythos research suggests that bug-finding capability is still improving quickly. The kernel's fix rate is racing against that curve. A week when fixes outpace reports is encouraging, but it is not a permanent settlement. The more useful question is whether the kernel can keep the ratio healthy as the tools get better.
For now, Linux 7.4-rc2 offers a piece of good news in the AI-and-software story. A major open-source project says AI-driven fixes made up much of a release candidate, and one subsystem is closing bugs faster than they arrive. The claim is limited, provisional and worth verifying. It is also an early sign that the patch flood might, eventually, be manageable.
