What did Linus Torvalds say about AI fixes in Linux 7.4-rc2?
On 8 November 2026, Linus Torvalds published the second release candidate for Linux 7.4 to the kernel mailing list. Release candidates are test builds published weekly while a new kernel stabilises. An rc2 usually arrives with a modest pile of corrections. This one did not. Torvalds opened his note by calling it "a big -rc2 to go with the largest ever -rc1". He added that the diffstat nonetheless looked "fairly benign".
Part of that size is bookkeeping rather than risk. Torvalds explained that the diff was dominated by automated Coccinelle updates. It was a batch of scripted changes he had deferred from the merge window. The updates are mostly one- and two-line edits, and they spread across thousands of files. He said they dominate the statistics because they are so spread out and add "a lot of noise".
Once that noise is masked out, Torvalds said, the shape of the release looks ordinary. About a third is drivers, with graphics, networking and sound accounting for most of it. Another third covers networking and BPF, and the remainder is spread across filesystems, memory management and architecture code. Nothing in that breakdown suggests a destabilised kernel; it suggests a large, busy one.
Then came the line that travelled. "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. That short sentence is arguably more significant than the release itself. AI-driven corrections have moved from a novelty worth flagging each cycle to an assumed part of how the kernel is maintained.
The specific signal followed in the next paragraph. Alexei Starovoitov is a co-creator and lead developer of eBPF. He had told Torvalds that fixes on the BPF side now arrive faster than new bug reports. Torvalds called that "lovely if true". He then tempered it immediately. Most other areas have not reached that point. The kernel may only be starting to see "the tide starting to turn".
The tone of the note is worth reading closely. Torvalds spends most of it on ordinary release engineering. He explains which trees contributed changes, why the scripted updates were delayed, and what testers should do next. The AI discussion takes two short paragraphs. That proportion is itself a statement. A subject that dominated kernel security arguments earlier in 2026 now sits inside a routine weekly update. It is discussed in the same register as driver fixes.
Why is the Linux kernel seeing so many AI-driven fixes?
The flood is not unique to Linux. In April 2026, Anthropic launched Project Glasswing. It is a collaboration with roughly 50 partners. The partners scanned important software using an early snapshot of Claude Mythos Preview. In a May update, Anthropic said the partners had collectively found more than ten thousand high- or critical-severity vulnerabilities. Several partners reported that their bug-finding rate had risen more than tenfold.
Anthropic also scanned open-source code directly. It said Mythos Preview had examined more than 1,000 projects and estimated 6,202 high- or critical-severity bugs among them. Independent firms assessed 1,752 of those findings; 90.6% proved to be valid, and 62.4% were confirmed as high or critical. One example was a certificate-forging flaw in the wolfSSL cryptography library, tracked as CVE-2026-5194, which Anthropic says is now patched.
The bottleneck moved quickly from discovery to repair. Anthropic wrote that a high- or critical-severity bug takes about two weeks to patch on average. Some maintainers asked the company to slow its disclosures because they needed more time to design patches. By May, Anthropic had reported 530 high- or critical-severity bugs; 75 were patched and 65 had public advisories. It had also sent 1,129 unvetted reports at maintainers' request.
Anthropic's public disclosure dashboard, last updated on 2 November 2026, shows the scale more clearly. It records 5,087 vulnerabilities disclosed across 680 open-source projects, with 629 patched to Anthropic's knowledge. Behind those reports sit 37,966 candidate findings and 6,376 triaged by external firms, of which 94.2% were confirmed valid. The dashboard also lists high-severity use-after-free findings in the Linux kernel repository itself.
Commercial software shows the same pressure. Microsoft fixed 974 vulnerabilities in its September 2026 Patch Tuesday release. The total included two exploited zero-days and 20 flaws rated potentially wormable, according to SecurityWeek. Rapid7 counted 723 of those flaws in Windows and called it the most CVEs Microsoft had ever published in a single day. Anthropic says Microsoft has warned that patch volumes will keep trending larger for some time.
Open source absorbs that pressure differently. A company can hire more security engineers; a volunteer maintainer often cannot. Anthropic acknowledged the problem directly in its May update. It noted that maintainers already face "a deluge of low-quality, AI-generated bug reports". That burden lands on top of their normal work. Its own process is designed around that constraint, with external firms reproducing findings and writing reports before anything reaches a maintainer. Even so, the company says it sometimes discloses directly when a project asks it to.
What is eBPF, and why does the fix rate there matter?
eBPF is a mechanism that lets small, sandboxed programs run inside the Linux kernel in response to events. Networking tools use it to filter and route packets. Observability platforms use it to trace system activity, and security products use it to enforce policy. Because those programs execute in kernel space, mistakes can affect the whole machine. That makes eBPF both a popular research target and a subsystem where fixes matter to a lot of production systems.
A fix rate that exceeds the report rate is a different measure from a raw patch count. If maintainers close more verified bugs than arrive each week, the backlog shrinks. If reports keep outrunning fixes, the queue grows even when everyone works flat out. That distinction is why Starovoitov's observation matters. It describes a crossover, not simply a busy week.
eBPF is also a plausible place for that crossover to appear first. It has dedicated maintainers, a large test suite and a steady flow of tooling work. Its patches are reviewed by people who know the subsystem well. Its users include cloud providers with a direct interest in stability. A well-staffed subsystem can absorb a surge that would swamp a smaller corner of the kernel.
The limits are just as important. No counts accompanied the claim in Torvalds' post, and the kernel does not publish a weekly ledger of reports against fixes. The statement is second-hand, covers one subsystem, and says nothing about severity. Closing more reports than arrive could still leave serious bugs open, particularly if the easy findings are fixed first.
The practical effect depends on where eBPF is deployed. Cloud networks, container platforms and security agents load BPF programs across large fleets. A fix merged upstream reaches those products through distributions and vendor kernels, not instantly. That makes the subsystem unusual. A small change can affect millions of machines, but only after several layers of release engineering.
How big was the Linux 7.4 merge window?
The merge window is the two-week period after a kernel opens for new work, when maintainers send large feature branches to Torvalds. Linux 7.4 took in almost 20,000 commits, against roughly 15,000 in a typical cycle. Torvalds gave that figure in his first release candidate post on 1 November. LWN counted 6,753 non-merge changesets by 23 October, while noting that the window might run to three weeks.
Torvalds framed the volume as a change for the better. He wrote that he was much happier seeing AI-triggered work during the merge window. The alternative was a stream of extra fixes during the stabilisation period. He called that approach "a much healthier model", because features and fixes land while maintainers are still actively reviewing their subsystems. He also cautioned that people expect 7.4 to be the next long-term-support kernel, which is an independent reason for a larger release.
Not all of the churn is meaningful. The Register noted that AMD GPU register offsets and bit masks made up about a third of the changes. Torvalds' reaction was dry: the project has seen this pattern before and will see it again. "Just ignore the noise," he wrote. Recurring register churn inflates line counts without changing behaviour, which is why diffstat size alone is a poor guide to risk.
The release also contains ordinary engineering. Phoronix reported that rc2 included a FAT32 optimisation it measured at up to 162 times faster. The release also brought a batch of sound-driver fixes and device quirks. AMDGPU also received corrections for HDMI 2.2 fixed-rate link support. Those items would have been the story in a normal cycle. In this one, they sit beside a security debate about how much AI has changed kernel maintenance.
How strong is the evidence that the AI patch flood has peaked?
The headline claim is stronger than the underlying text. The Register reported the rc2 post under a headline saying Torvalds "sees signs" the AI-prompted patch flood "has peaked". Its opening line went further, saying the worst of the "patchpocalypse" may be behind us. Torvalds' own text is more cautious. He said one subsystem appears to have turned a corner. Most others have not, and readers should cross their fingers rather than hold their breath. The difference matters for anyone planning security work around the story.
Patch counts are a noisy proxy in both directions. Coccinelle batches and register churn inflate totals without adding bugs, while a single fix can close several duplicate reports. A report-versus-fix ratio is better, but it still measures activity rather than residual risk. Nor does it say whether the fixes are correct. Kernel maintainers rewrite submissions routinely, and Torvalds has said AI-suggested patches often get rewritten before they are merged.
The quality data that does exist comes from Anthropic's triage, not from the kernel. Its 94.2% true-positive rate covers 6,376 findings that external firms reviewed. That is a strong result for a bug-finding pipeline, but it is a different population from patches submitted to Linux. It also implies that roughly one in 17 reviewed findings was not a valid vulnerability. At the scale of tens of thousands of candidates, that rate becomes expensive.
Capacity remains the binding constraint. Anthropic reported that some maintainers asked it to slow down. By May, only 75 of 530 disclosed high- or critical-severity bugs had been patched. Patching is also not the same as deploying: a fix must reach distributions, containers and long-term-support branches before most users are protected. That tail is measured in months or years, not release candidates.
A peak in the flow is also unlikely to be permanent. A future model could repeat the jump that Mythos Preview produced in April. If it does, report volumes will rise again regardless of how well maintainers handled this cycle. Torvalds' observation describes a moment in a moving system, not an endpoint.
Severity adds another layer of uncertainty. A count of reports treats a minor null-pointer check and a remotely reachable memory-corruption bug as equivalent. Anthropic's own data shows why that matters: only 62.4% of its assessed high- or critical-rated findings were confirmed at that severity. A subsystem could clear a flood of low-impact reports while serious bugs remain queued behind them.
What should Linux users and maintainers watch next?
The next few release candidates will test the claim. If the report-versus-fix crossover holds, later rc diffstats should show fewer AI-driven corrections in BPF and, eventually, in neighbouring subsystems. If the effect was a one-week artefact, the correction volume should return. Torvalds' weekly notes are the cheapest place to watch, and Starovoitov's own pull requests would provide harder numbers than a second-hand summary.
Linux 7.4 is widely expected to become the next long-term-support kernel. That raises the stakes for the fix flow. LTS branches receive backports for years and are the base for many enterprise distributions. A larger-than-usual set of security corrections in 7.4 will translate into larger stable updates later. Administrators should plan for that rather than assume the current cycle is a temporary spike.
Anthropic's dashboard is the other public gauge. It stood at 5,087 disclosed vulnerabilities on 2 November, and its patched count will show whether the repair side is catching up. Other model providers are building similar capability, so the supply of findings is unlikely to fall. The question is whether maintainers gain tooling and review capacity at a comparable pace.
For users, the practical advice has not changed. Systems on the 7.3 stable series should keep applying security updates. Anyone testing 7.4-rc2 should report regressions through the usual kernel channels. The durable change is not a single hopeful sentence from Torvalds. It is that AI-assisted finding and fixing is now treated as normal kernel work, with all the review burden that implies.