Posted by flipped 15 hours ago
Which illustrates pretty well something that's lost when relying heavily on LLMs to do work for you: exploration.
I find that doing vulnerability research using AI really hinders my creativity. When your workflow consists of asking questions and getting answers immediately, you don't get to see what's nearby. It's like a genie - you get exactly what you asked for and nothing more.
The researcher who discovered Copy Fail relied heavily on AI after noticing something fishy. If he had to manually wade through lots of code by himself, he would have many more chances to spot these twin bugs.
At the same time, I'm pretty sure that by using slightly less directed prompting, a frontier LLM would found these bugs for him too.
It's a very unusual case of negative synergy, where working together hurt performance.
Very much aligns with my experience. For me this is the most unsatisfying thing about AI-based workflows in general, they miss stuff humans would never miss.
All the time I wonder what am I missing that's right nearby? It's remarkable how many times I have to ask Claude code to fully ingest something before it actually puts it into context. It always tries to laser through to target it's looking for, which is often not what you want it to look for, at least not all you want it to look for. Getting these models to open up their field of vision is tough.
I have found that the “pro” approach is much more holistic and able to tackle rather “creative” problems that require very careful design and the overall artifact is tight and self-consistent. — Claude Code by comparison is incredible in exploration and targeted implementation but indeed is not great at seeing the forest.
But what if a coding agent was prompted to be more curious during development? Like a human developer, make mental notes of alternatives to try out and chase suspicious looking code which may seem unrelated to the task at hand. It could even spawn rabbit hole agents in parallel.
Taking a step back, this probably highlights major hazard with the increased usage of LLMs for coding, which is that everyone's style of work is going to converge because most code will be written by the 2-3 most popular models using the same system prompts.
I've had to explicitly direct the machine to read existing sibling code and follow the specific idioms and patterns in use.
> All the time I wonder what am I missing that's right nearby?
The wrong thing got fixed for copy.fail, because people jumped to blame AF_ALG.
[ed.: yes it's the same authencesn issue. https://github.com/V4bel/dirtyfrag/blob/892d9a31d391b7f0fccb... it doesn't say authencesn in the code, only in a comment, but nonetheless, same issue.]
[ed.2: the RxRPC issue is separate, this is about the ESP one]
The RxRPC one is definitely a different root cause (although caused by a very similar mistake).
For the ESP one it's a bit harder to tell. I don't think the wrong thing was fixed, just that there was a very similar bug in almost the same spot. Could be wrong about that though.
It's absolutely the same issue in authencesn/ESP. There's another one in RxRPC that is AIUI completely unrelated.
I bet that with a slightly looser prompt/harness, the LLM could have found these twin bugs too.
Yet at the same time, I also think that if the human researcher had manually scanned the code, he'd have noticed these bugs too.
FWIW I do think LLMs are great tools for finding vulnerabilities in general. Just that they were visibly not optimally applied in this case.
I think LLMs are great for vulnerability discovery, but you need to not skimp on the legwork and understanding what even you just found there.
They do not get bored like a human but they are trained on human language and replicate the same traits, such as laziness, and expressing boredom or annoyance (even if obviously they do not experience anything at all). It’s actually a lot of effort to get them to engage with things at a deeper level without skipping corners
> This finding was AI-assisted, but began with an insight from Theori researcher Taeyang Lee, who was studying how the Linux crypto subsystem interacts with page-cache-backed data.
There's no question that we live in the world where LLM AI was involved in finding the copy fail vulnerability at this specific time, and it's completely normal for people to see a vulnerability and then look closer and find related vulnerabilities or a deeper root cause, but there's no need to adopt an extreme "without AI LLM we don't find these vulnerabilities" position.
That is a very difficult fact pattern to which to attach the conclusion "LLMs have sabotaged security research" (my paraphrase).
Also I see you jumping around a lot to the defense of LLMs when I don’t think anyone is really attacking them. Maybe cool it a bit.
So like I said, just chill out.
Is there a counterfactual where you would say it explored well enough, besides both vulnerabilities published as one?
If there's a root cronjob that runs a world readable binary, you could modify it in the page cache and exploit it that way.
Modifying the page cache is a really strong primitive with countless ways to exploit it.
So it seems surprising to me that you can call it when the out fd is not writable? But I didn't retain the information about the vulnerability, so I'm missing something. There was something about copy on write, IIRC?
Also, NB, I said permission check, not mode check. The input fd to splice can and will be open for only reading quite often. Doesn't mean the kernel can't still do a write permission check.
(Except I didn't say that here. Oops. Getting confused with my posts.)
But splice is a more or less a generalization of sendfile, and sendfile is often used for webserving where the serving process does not have ownership of the documents it is serving. It doesn't make sense to limit splice such that it can't do the task it was built for. Maybe splice should just not write to the input fd? :P
Not really, splice(2) is actually more limited, it's an optimisation for reading and writing data between files and pipes without needing to make copies.
sendfile(2) works with any fds because it just exists to remove a fair bit of the copy overhead when doing a userspace read/write loop, but it does actually do a copy.
But apparently we can't be trusted with the page cache…
Maybe the kernel using supervisor-read-only flags could be made to work, only issue then is what happens if something does in fact need to write…
That's why is very very important to just step out and use saved time to go for a walk, to a park, sit on a bench, listen do birds, close eyes and zoom out.
The state we are in is actually brilliant.
link: https://github.com/V4bel/dirtyfrag
detailed writeup: https://github.com/V4bel/dirtyfrag/blob/master/assets/write-...
importantly:
"Copy Fail was the motivation for starting this research. In particular, xfrm-ESP Page-Cache Write in the Dirty Frag vulnerability chain shares the same sink as Copy Fail. However, it is triggered regardless of whether the algif_aead module is available. In other words, even on systems where the publicly known Copy Fail mitigation (algif_aead blacklist) is applied, your Linux is still vulnerable to Dirty Frag."
mitigation (i have not tested or verified!):
"Because the responsible disclosure schedule and the embargo have been broken, no patch exists for any distribution. Use the following command to remove the modules in which the vulnerabilities occur."
sh -c "printf 'install esp4 /bin/false\ninstall esp6 /bin/false\ninstall rxrpc /bin/false\n' > /etc/modprobe.d/dirtyfrag.conf; rmmod esp4 esp6 rxrpc 2>/dev/null; true"
sudo echo 3 > /prox/sys/vm/drop_cachesAnd if a machine is already exploited, it's too late to do just that. You need to rebuild the whole disk image because anything on it could be compromised.
this is more targeted at the people who run the PoC to see if their machine is vulnerable.
just transcribing some relevant stuff from https://github.com/V4bel/dirtyfrag/issues/1 so that people visiting this thread dont need to poke around a bunch of different places.
> 2026-05-07: Submitted detailed information about the vulnerability and the exploit to the linux-distros mailing list. The embargo was set to 5 days, with an agreement that if a third party publishes the exploit on the internet during the embargo period, the Dirty Frag exploit would be published publicly.
> 2026-05-07: Detailed information and the exploit for this vulnerability were published publicly by an unrelated third party, breaking the embargo.
Edit: nevermind, details are further down in the thread:
https://openwall.com/lists/oss-security/2026/05/07/12
And
> on 2026-05-05 Steffen Klassert pushed f4c50a4034 to netdev/net.git with Cc: stable@vger.kernel.org.
Once a fix is out it's usual for researchers to race to make the first exploit out of it.
echo 3 | sudo tee /proc/sys/vm/drop_caches
sudo sh -c 'echo 3 > /proc/sys/vm/drop_caches'
sudo sysctl -w vm.drop_caches=3 su -c 'echo 3 > /proc/sys/vm/drop_caches'Tested locally on Ubuntu 26.04:
1. Ran the exploit and got root
2. Configured the mitigations
3. Ran `su` again with no parameters and immediately got root again unprompted
4. Cleared the page cache
5. `su` asked for a password
authencesn didn't get fixed. Now we got the results of that, turns out you can access the same (I believe) out of bounds write through plain network sockets.
I wish I thought of that, but I didn't.
[ed.: I'm referring to the through-ESP issue. The RxRPC one is AIUI completely unrelated.]
https://lore.kernel.org/lkml/2026050851-iron-hurdle-6421@gre...
https://lore.kernel.org/lkml/2026050843-unplowed-spinster-cf...
https://lore.kernel.org/lkml/2026050832-remold-faceless-bed0...
https://lore.kernel.org/lkml/2026050825-heaving-spender-13a8...
This feels like the practice of Linux distros back in 1999 when they'd ship default installs with dozens of network services exposed to the internet. Except it's not 1999 anymore.
And... I remember the early days of Linux where I ran `make menuconfig` and selected exactly the functionality I wanted in my kernel. I'd... rather not end up back there.
That said a target for an easy win here is RHEL, which compiles a lot of modules into the kernel rather than leaving them as loadable modules, so the mitigation for e.g. copy fail was impossible. Maybe they could do with a few less of those?
Well, because you probably don't, and it's a security risk, so no need to put millions at risk for the benefit of that one person who wants to tinker with packet radio or whatever. Similarly, it would be prudent for distros to not allow autoloading of modules that are extremely niche while giving a simple way to adjust the settings if you want to. God knows they have plenty of GUI configurators and config files already.
No reason why you couldn’t just `dnf install -y kmod-rxrpc` if for whatever reason you need that.
If the kernel modules for esp4, esp6 and rxrpc aren't loaded - how is it that a non-root attacker can cause them to get loaded?
From the sound of it, the same mitigations for Copy Fail 1 are also effective here.
We have forgotten what a distro is, and its modern corruption of the concept is now taken as the definition.
Distributions weren't meant to be competing generic universal bundles of userspace tools in addition to the kernel.
Linux distro maintainers are the most responsible software maintainers on the planet. Their security practices are miles beyond the stupid programming language package managers, they maintain a select list of packages, vet changes, patch bugs, resolve complex packaging issues, backport fixes, use tiered releases, distribute files to global mirrors, and cryptographically validate all files. And might I remind you, they do all this for free.
Today it's 0.1%, tomorrow it might become 100%. User demand is hard to anticipate, so it's reasonable to include small features that don't cost a lot to run by default.
It's not ideal, but you really don't want to prevent user from finishing their task, because maybe then they'll just give you a bad name and switch to another distro.
That's to say, it's not "irresponsible", it's reasonably maximums (at least trying to be).
https://durovscode.com/google-android-security-update-warnin...
The claim is Android is much more secure than other Linux, but if 40% of all Android devices don‘t get a security patch and you can’t even do it yourself I would call the more secure per se.
Hardening is one part of security, patchability another. Android lacks in the latter.
(Of course the problem isn't Android, it's the chipset vendors that the SW depends on. They drop support fast and never give enough info for anyone else to keep things up to date. Also Google.)
No system will stay secure once it does not receive updates. That does not exclude it from being more secure than another system based on security feature merits as long as it does get updated.
>Hardening is one part of security, patchability another. Android lacks in the latter.
That is not an inherent flaw with android but OEM devices shipping modified android they don't bother keeping up to date. Some OEMs are trying to mitigate this by increasing security update support up to 7 years which still is not long enough but also doesn't make them less secure than a desktop that gets updated longer.
What people forget is that not only desktop and mobile phone software is different but also the hardware. If your desktop pc hardware is out of date / EOL nobody cares usually. Meanwhile on a phone this can be a lot more relevant because security expectations and threat models are a lot higher, for example see all the zero/one click compromise headlines.
Having a service that automatically starts and listens on the network is radically different from having a module that a local administrator can load.
If you want to block module loads, you’re one sysctl flag away.
This is a successful local privilege escalation, so local administrator privs were not needed. In default configuration of all distros, apparently.
> If you want to block module loads, you’re one sysctl flag away.
The modules aren't really the point, it's that unnecessary features (to 99% of us?) were accessible by default without privs.
It's not any different from putting an always-running network service behind socket activation instead. The security boundary/risk is nearly identical between the two.
You responded contrasting a network service with an administrator-loadable module.
This is neither of those. It's an LPE, not a remote exploit. It doesn't require an administrator (root) to load anything. In context of this vuln, it's exactly analogous to socket activation. The scope of an LPE vuln is local; yes. What does that have to do with the rest of your comments?
I originally replied to a comment saying "This feels like the practice of Linux distros back in 1999 when they'd ship default installs with dozens of network services exposed to the internet". It is not like that.
Par for the course for HN.
This would reduce the amount of ring 0 code. But I've never seen such advice.
If an attacker manages to do all that, its already bad news for you. Escalation to root with this is the least of your worries at that point.
Like someone else below posted, https://xkcd.com/1200/
People need to understand what the vulnerability actually is before freaking out about it.
Hell, GitHub Actions would do.
We could also wonder why XZ was linked to SSH... But only on systemd-enabled distros (which is a lot of them).
Just... Why?
And then make sure to call to incompetence, instead of malice and say non-sense like "Sure, it only factually affects systemd distros, but this is totally not related to systemd". All I saw though was a systemd backdoor (sorry, exploit).
Now regarding copy.fail that just happened: not all maintainers are irresponsible. And some have, rightfully, bragged that the security measures they preemptively took in their distros made them non vulnerable.
But yup I agree it's madness. Just why. And Ubuntu is a really bad offender: it's as if they did a "yes | .." pipe to configure every single modules as an include directly in the kernel.
"We take security seriously, look we've got the IPsec backdoor (sorry, exploit) modules directly in the kernel". "There's 'sec' in 'IPsec', so we're backdoored (sorry, secure)".
Or, y'know, offer some forms of compute as a service.
Native unsandboxed execution == root. Only thing that's new is some people started making websites for their LPEs.
https://github.com/google/security-research/tree/master/pocs...
Within an hour of be advised of, and running the mitigation for DirtyFrag, my upstream provider has blocked all WHM/cPanel/SSH/FTP/SFTP access with a heads-up on:
CVE-2026-29201 CVE-2026-29202 CVE-2026-29203
which look like a repeat of CVE-2026-41940 a week ago.
Physical access always means the device is pwned. You can install a keylogger or something similar.
I tried fixing the paths and even linking `/bin/bash` to the nix /run/current-system/sw/bin/bash
/etc/passwd is unmodified.
Can anyone else try? CopyFail1 did not work because `su` is only executable, not readable, CopyFail2 worked only partially (changes /etc/passwd but the user is not passwordless)
Linux is open source, so every patch fixing the security bug is immediately visible to everyone. There is no workaround to that by the very design how the kernel is developed. The "embargo" people talking about is the rather stupid notion that if people keep their mouth shut and not write "THIS IS A LPE" straight in the patch description, everyone can pretend vulnerability is not leaked until the "official" message in the mailing list is sent.
This approach might have been defensible before, but in LLM era, when people have automated pipelines feeding diffs straight from the mailing lists to SotA models asking to identify probable security issues fixed by those, it is both stupid and dangerous.
For Linux/public open source, what you said is right about 2). Once the patch is visible to anyone, it's trivial to identify exploits for unpatched systems. But 1) is still a valid use-case for embargoes for Linux vulns, right? Like, if this patch had taken a few weeks to develop before being confirmed working and published, that's potentially valid grounds for not sharing details during that time (within reason), no?
Usually, nobody even bothers to check. LPEs like this are too common to even categorise effectively.
https://xcancel.com/encrypted_past/status/205240982299839296... https://xcancel.com/encrypted_past https://github.com/0xdeadbeefnetwork https://github.com/0xdeadbeefnetwork/Copy_Fail2-Electric_Boo...
It sounds like these two most recent exploits depend on unprivileged user namespaces, and that in fact a high percentage of LPE exploits need this feature. I use rootless containers on a couple of systems (like my dev machine server), but on most of my systems I don't, so it sounds like disabling that would be a good step to hardening my systems against future exploits.
To the security experts: are there any other straightforward configuration changes with such broad-reaching improvement in security posture? Any well-written guides on this subject, something like "top kernel modules to consider disabling if you don't need them"? I'm not talking about the obvious stuff like "disable password SSH", I'm specifically looking for steps that are statistically likely to prevent as-yet-unknown privilege escalation attacks.
I don't have any guides but you can determine which kernel modules are already loaded in your system and then just compile those in and block module loading.
Otherwise, shove everything into a container, ideally gvisor, and you've reduced attack surface by a large chunk again via seccomp.
For anyone not on the security stream of Debian packages for Bookworm, kernel version 6.1.0-42-amd64 is actually immune to copy.fail. Surprising that it looks to be immune to dirtyfrag. If you haven't already patched on the security stream, you can choose any kernel version that kept commit 2b8bbc64b5c2. I am thinking that the same commit might accidentally be keeping certain Debian 12 kernel versions safe from dirtyfrag as well.