Something good coming out of AI, it seems.
You must log in or register to comment.
People here can’t seam to get past there built in kneejerk reaction to anything AI related.
oh wow fam, ai is protecting us from… ai
Or from NSA backdoors.
How many of them are new?
Most likely very few of them because the code still gets reviewed by humans.
This is misleading.
Edit: “misleading” was the wrong word. I was just pointing out that showing data from 4.x to 5.x, 5.x to 6.x also would have made this post more meaningful.
Linux kernel full releases always have CVEs shortly after them because new features, drivers, memory handling tuning, and other capabilities arr introduced, and people (or ai) hammer away at them as soon as they appear.
I acknowledge that the bump has more amplitude because ai tools accelerate this, but this graph isn’t the shocker you think it is.
I’d love to see the graph include data back to 4.x.
“Axis bias” is real
I thought Torvalds said major versions bumps were meaningless so 6.0 or 7.0 is the same as 6.12 to 6.13
Minor versions go up according to a certain pattern based on how the change mgmt is handled for the kernel, but major releases are marked when the kernel team “feels it’s the right time”.
From lwn: https://lwn.net/Articles/871989/
To kernel developers, version numbers continued to have no meaning other than as a measure of where things were on the timeline, he said; users were encouraged to run the latest version. But if you tell customers they need to move from 3.20 to 4.0, they will consider it a multi-year effort that requires planning and paperwork; “so they will only do it when they have time, which is never”. Going from 3.20 to 4.0 seems like a big jump, and 4.0 to 4.2 seems much smaller, even though the latter brings roughly double the number of changes.
Maybe in the past, but that doesn’t seem to be the case at the moment.
Torvalds pointed out that the numbers he applies to new kernel releases are essentially meaningless.
“We haven’t done releases based on features (or on “stable vs unstable”) for a long, long time now. So that new major number does not mean that we have some big new exciting feature, or that we’re somehow leaving old interfaces behind. It’s the usual “solid progress” marker, nothing more.”
Thanks for this, I think I now understand what others are sating about this.
The version numbers are not really meaningless, because anything higher than .20 is life-threatening to Torvalds. But technically and functionally, it doesn’t matter.
Lol he said he doesn’t want any version number higher than what he can count on his fingers and toes.
What does he do when he gets to 19.19?
I hope he is still around at that time.
This is misleading.
Is it? What exactly is it misleading here? It just aligns with the usage of Ai, which means that Ai tools helped finding and fixing those issues. I don’t see anything misleading.
The meaningful data would have been to show the amplitude of the commit frequency bump compared to 5.0 and 6.0 releases, because, as I mentioned, they have bumps just after release also.
This just shows a bump after a release, which doesn’t really show “AI” anything.
It shows that the CVE fixes were consistent during the last 2.5 years and 14 releases. And with the recent rise of Ai usage in the Kernel, this bump aligns.
commit frequency bump compared to 5.0 and 6.0 releases
Why 5.0 and 6.0, those are ancient. This goes back to 6.9 from May 2024, where Ai usage in the Kernel wasn’t a thing until the recent few releases.
It’s not AI usage in the kernel - it’s the increase in AI to do code auditing that had seen the spike in bug reports. A lot of FLOSS projects are seeing this regardless of their use of not of AI to generate code.
It shows that the CVE fixes were consistent during the last 2.5 years and 14 releases.
Right, yes. But if the transitions from 4.x to 5.0 and 5.x to 6.0 were also shown, we’d be able to see that the bump itself isn’t significant, and that this happens after every major version release.
The magnitude of the bump is what’s important, and we can’t see that without comparing to pre- and post- AI tooling.
Why 5.0 and 6.0, those are ancient. This goes back to 6.9 from May 2024, where Ai usage in the Kernel wasn’t a thing until the recent few releases.
That’s the whole point.
There is no need to go older. We have 14 releases in between 6.9 and 7.2.
I don’t understand what you’re saying… Those aren’t major releases, those are dot releases, and they cross a major release.
Linux versioning does not work like that. There is no major release, the are all equivalent. It’s not a joke, Linus Torvalds just don’t want high numbers, so at the latest .20 (dot-twenty) release they create a new major one. But functionally it doesn’t matter. Therefore the release version is meaningless for our discussion. I’m really serious about this, this is not in any way trolling or being toxic. All we need for this discussion are already present in the graph.
Have a look at: https://en.wikipedia.org/wiki/Linux_kernel_version_history
In contrast to semantic versioning, the major version carries no intrinsic meaning for the kernel, as Linux promises stable interfaces and behavior even when the major version changes. However, since version 3.x, minor versions have been capped around 20 to avoid the false perception that changes between large minor versions (e.g. from 5.30 to 5.31) would be less voluminous or impactful than changes between smaller minor versions (e.g. from 6.1 to 6.2).
I think the other persons point is that if you were to start at 6.0 and sum all the CVEs until 7.0, the numbers would be similar to the CVEs from 7.0
This doesn’t make sense. We are comparing single releases, not the entirety of 6.0 up to 7.0. How many years would this be? Also why do people want see 6.0? We have 14 last releases, in span of 2.5 years. And only the last two or 3 releases got a huuuge spike, which coincides with the broad usage and advancements of Ai in the Kernel.
We don’t need to compare older versions. The data is here and enough to support this claim. These are 14 releases for gods sake.
The graph would’ve been stronger if other versions were included. Something like this is for example, (it’s lowk a bad example though since i can’t really tell much about what the author is claiming as an effect of the figure)
https://www.blogspan.net/linux-kernel-cve-2026-98432-tcp-ubuntu-fix/
Which shows compared to previous years a similarish rise in cves but they appear more trivial (for a variety of reasons such as enhancing requitements to be high risk cve)? Would be worth replication
The graph would’ve been stronger if other versions were included.
It shows versions up to v6.9 from May 2024, which is a span of 2.5 years approx. and 14 releases. And with the knowledge that Ai usage accelerated during the last few releases, we can conclude its a major factor here for the rise.
How is it misleading? All you’ve said is that releases have cves which is what the graph shows.
As I mentioned, all kernel major releases have cve bumps, because (also mentioned) new kernel releases have new features and drivers.
The bump itself doesn’t show that ai is associated with more CVEs, because that bump is present after every kernel major release. It would have been better to show the bumps from previous releases to demonstrate the volume ai tools are helping to find.
That’s not how the version numbers for the Linux kernel work. 7.0 is the release that follows 6.19, they are both new mainline releases in the exact same way, and the first number increasing doesn’t mean anything.
Yes, acknowledged. I see the point some ppl were trying to make.
By “bump” you mean the quadrupling of CVEs being addressed?
Or being introduced. Or neither of those two, because it’s an isolated graph devoid of larger context
It’s “being discovered”. You need context? Here: https://zerodayclock.com/#where-the-vulnerability-exposure-is-growing
Yes, this is the point I was trying to make, there is no context.
But we don’t have enough information here; it says “CVE fixes per release”, we don’t know anything about them. Are these long-standing? Is some new process introducing security issues? We also have ongoing conversions of modules to rust, and lots of ongoing work.
We can say AI tools are helping find these, and that may be true, but but this graph doesn’t say that, it just says “CVE Number Big”.
Something good coming out of AI, it seems.
Yeah, wasn’t the security community just complaining about the number of CVEs published recently which were just unfiltered bug reports?
The more fixed, the better.
how many are introduced by new ai code?
Assuming that Linux kernel team continues to use the same review process they always had, there’s zero reason to think that there would be an increased rates of bugs due to LLM generated code. This seems to be something people without programming experience don’t understand. Code quality isn’t premised on trust, it’s a product of the development process. Projects ensure quality by having specifications, tests, doing code reviews, and so on. As long as the practices the kernel team has established continue to be applied, they will catch the bugs made by LLMs or humans exactly the same way.
Idk man. I’ve certainly become more lenient in Code Reviews because I have to do like 5 a day instead of 2-3 a week, all with like 2k lines of changes. I literally don’t have the time to fully read tests anymore, I mostly just skim them and check whether they have an assess statement that looks somewhat reasonable.
Hell, I’ve recently had a meatpuppet of a coworker brag about how his implementation of a Card component for the UI library had 199 tests. I’m so glad I wasn’t assigned to review that shit. But it sure looks great to the higher-ups 🫠
Hopefully working on the Linux Kernel doesn’t suck as much as working for a big corp.
Large corps always end up implementing shitty practices because business practices are driven by pencil pushers. Even before LLMs, some of the most soul crushing jobs were found at large corps working with shit like J2EE. That’s nothing new really.
At my job, we use LLMs, but nobody is making 2k loc PRs all of a sudden, and we still pair up to review code, and generally work at human speed.
As long as the practices the kernel team has established continue to be applied,
Right, the question is, if they’re changing the writing code practices, do you think there might be any change in the reviewing process?
I’ve seen no evidence of that so far. I also fail to see how writing code practices are related to the review and quality control process. These things aren’t linked in any way that I’m aware of.
The ability to comprehend code underlies both of those and is trained by both. Doing less of one reduces your training.
LLM use leads to skill deficits, it’s shown up in studies but also just, use it or lose it.
The only studies I’m aware of used tiny samples and produced muddled conclusions. If you have some conclusive empirical research then do feel free to share it though. Meanwhile, using LLMs doesn’t mean doing less, it just means focusing on different things. The skill deficit here is the same kind as the skill deficit people have using punch cards or writing assembly by hand today. As things become automated, different things become relevant and that’s what people shift their attention to.
If this is primarily fixing old issues, over time we should see a sharp drop off and then even fewer security fixes then before due to fewer being introduced, due to being caught by AI upfront.
On the other hand, if AI is introducing more issues than it is resolving, we’ll see the baseline of security issues being found permanently addressed.
