This is very similar in root cause and exploitation to Copy Fail.
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.
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.
Or a follow up prompt: "find similar classes of bugs". Once the actual case has been layed out finding like bugs isn't too hard. I hear you on the creativity bit. Like any tool, AI can put blinders on. Using it to augment without it fully taking over your workflow is tough.
I don't follow. LLMs spotted these bugs in the first place. You seem to be saying that these discoveries are indications that they're bad for vulnerability discovery.
I don't think the copy.fail people understood the issue they found, as is evident by the heavy focus on AF_ALG/aead_algif, which is essentially "innocent" as we're seeing here.
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.
From what I understand, the copy fail bug was found by researcher who noticed something weird and then using AI to scan the codebase for instances where that becomes a problem.
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.
No, they did not. Careful of falling for the psychosis.
> 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.
These are all page cache poisoning attacks (dirtyfrag, copyfail, dirtypipe). Maybe the page cache should have defense-in-depth measures for SUID binaries?
True! Building protections (e.g. physical pages in the page cache are not writeable 100% of the time) just for executables has of course countless circumventions as well (e.g. config files). Yeah, there is probably not that much to be done there, actually. Looking at some of the diffs it seems to me like the kernel makes it really not particularly obvious when/how this goes wrong. E.g. the patch for this is to look at an additional flag on the socket buffer to fix an arbitrary page cache write. This feels rather action at a distance. Logically this of course makes sense, the whole point of splice et al is to feed data from one file-like into another file-like, whichever those ends might be. That erases the underlying provenance of the data.
"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."
"sudo" in "sudo echo 3 > /prox/sys/vm/drop_caches" does not do anything because only runs echo, not the write.
And 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.
>And 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.
If this indeed works on all major distributions, I just continue to be amazed by how irresponsible the maintainers are. We're talking about optional kernel functionality that's presumably useful to something like <0.1% of their userbase, but is enabled by default?... why?
This feels like the practice of Linux distros back in 1999 to ship with dozens of network services by default. Except it's not 1999 anymore.
And I ask again: why the f*ck is algif_aead getting all the flak for copy.fail? It's authencesn being stupid.
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.]
Every time someone finds a universal Linux privilege escalation, somewhere a sysadmin whispers 'this is why we don't run as root' while nervously checking if their containers are actually isolated.
I assume because the rxrpc module is not loaded / provided and because unprivileged user namespaces are not allowed, which should be sufficient to mitigate. Curious if someone else has more details though.
Yes, it demonstrates that it's possible to harden well - at least for some cases. It appears depending on the environment hardened kernel / runtime environments are pretty much possible to have safeguards working today already.
Because Android is not Linux, as much as some pretend it is.
In fact, given the official public APIs, Google could replace the Linux kernel with a BSD, and userspace wouldn't notice, other than rooted devices, and the OEMs themselves baking their Android distro.
It absolutely is Linux, and yes the JVM could absolutely run on something else. But it is Linux and you can run Linux binaries directly on it - that just isn’t how it is used by end users.
No you cannot, the NDK has a specific set of oficial APIS, and the Android team feels in the right to kill any application that doesn't follow the law of Android land.
Some folks like the termux rebels, occasionally find out there is a sherif in town.
> As documented in the Android N behavioral changes, to protect Android users and apps from unforeseen crashes, Android N will restrict which libraries your C/C++ code can link against at runtime. As a result, if your app uses any private symbols from platform libraries, you will need to update it to either use the public NDK APIs or to include its own copy of those libraries. Some libraries are public: the NDK exposes libandroid, libc, libcamera2ndk, libdl, libGLES, libjnigraphics, liblog, libm, libmediandk, libOpenMAXAL, libOpenSLES, libstdc++, libvulkan, and libz as part of the NDK API. Other libraries are private, and Android N only allows access to them for platform HALs, system daemons, and the like. If you aren’t sure whether your app uses private libraries, you can immediately check it for warnings on the N Developer Preview.
That's specific libraries, when using the default linker. You could construct that same behavior on desktop linux too. And you can avoid it equally well on Android - you can statically-link things just fine, you can use libraries you actually control, and presumably use a custom linker if desired. It's utterly non-surprising that "you run code you don't control" results in "said code...can do arbitrary things for unsupported use". (Never mind that, instead of a "sherif", they could've just renamed all private symbols, or just naturally replaced them over time, breaking your code all the same, just in a more confusing way)
Also some obligatory Linux vs GNU/Linux comment. (and it's not like GNU/Linux doesn't ever change under your feet - see the glibc DT_HASH debacle)
That's all user space platform specifics, it has no relation to your previous statement where you said 'android is not linux'.
Someone can statically build a freestanding executable/so targetting arm64 linux (specifically the right android linux kernel version) and it will run fine on Android. The syscall interface, process model, file descriptors, signals, memory mapping, all of this is Linux, this is what people mean when they say Android is just Linux.
They're asking the nature of the third party's discovery/publishing. Someone on the inside who decided to leak it anonymously? Someone else who was able to access some private communication they shouldn't have been able to see? Or a third party who happened to discover the same vulnerability (which seems less unlikely than normal since this is so similar to Copy Fail), but didn't follow disclosure procedures?
I think I read on the bug's website that "No fix has been released". I understood that as there is no public fix, but maybe it only means it's not in a tagged version of the kernel and no hotfixed distro kernels have been released?
2026-04-29: Submitted detailed information about the rxrpc vulnerability and a weaponized exploit that achieves root privileges on Ubuntu to [email protected].
2026-04-29: Submitted the patch for the rxrpc vulnerability to the netdev mailing list. Information about this issue was published publicly.
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 the esp vulnerability were published publicly by an unrelated third party, breaking the embargo.
2026-05-07: After obtaining agreement from distribution maintainers to fully disclose Dirty Frag, the entire Dirty Frag document was published.
With the exploits published as-is, you'll only get root inside the container: there's no explicit namespace break, and calling setuid() in a container just gives you root in the container.
However, it can be used to modify files that are passed into the container (e.g. Docker run -v), or files that are shared with other containers (e.g. other Docker containers sharing the same layers). kube-proxy with Kubernetes happens to share a trusted binary with containers by default, which is how it can be exploited: https://github.com/Percivalll/Copy-Fail-CVE-2026-31431-Kuber...
Do you think with modern LLMs in a few years projects like Linux will have all those low-hanging security bugs fixed? Are we witnessing a transition period, or will nothing change?
there's an argument to be made that new code will be inspected before being merged and therefore the classes of bugs an LLM is likely to find will not be merged until it's fixed.
Here's a general question, are these vulnerabilities hitting Linux more than BSDs due to hit being a larger target or because its architecture is less secure by design?
AFAIU, Linux and the BSDs have basically the same architecture - the BSDs just value secure and simple, understandable code more highly than Linux vs features and performance.
- more people are using it (assuming macos is in its own bucket perhaps)
- bigger surface areas (esp NetBSD has in my limited understanding just less stuff that can go boom)
- more churn, ie more new stuff than can be buggy released more often.
Of course, because of that, more eyes are on Linux, so I'm not sure where that security tradeoff is.
I agree with the general sentiment. I treat anything running arbitrary machine code as if it has full access to a machine. I don't know where you get "run your services as root" from that, though. The principle of least privilege doesn't just apply to running malicious code, but running buggy code whose attack surface is exposed to evil-doers.
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.
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 think LLMs are great for vulnerability discovery, but you need to not skimp on the legwork and understanding what even you just found there.
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.
> 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.
https://xint.io/blog/copy-fail-linux-distributions
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.
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."
conversation around the mitigation suggests you need a reboot or run this after the above on already-exploited machines:And 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.
This feels like the practice of Linux distros back in 1999 to ship with dozens of network services by default. Except it's not 1999 anymore.
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.]
- esp4 (kernel config "CONFIG_AF_RXRPC")
- esp6 (kernel config "CONFIG_INET_ESP")
- rxrpc (kernel config "CONFIG_INET6_ESP")
Is this correct?
In fact, given the official public APIs, Google could replace the Linux kernel with a BSD, and userspace wouldn't notice, other than rooted devices, and the OEMs themselves baking their Android distro.
Some folks like the termux rebels, occasionally find out there is a sherif in town.
> As documented in the Android N behavioral changes, to protect Android users and apps from unforeseen crashes, Android N will restrict which libraries your C/C++ code can link against at runtime. As a result, if your app uses any private symbols from platform libraries, you will need to update it to either use the public NDK APIs or to include its own copy of those libraries. Some libraries are public: the NDK exposes libandroid, libc, libcamera2ndk, libdl, libGLES, libjnigraphics, liblog, libm, libmediandk, libOpenMAXAL, libOpenSLES, libstdc++, libvulkan, and libz as part of the NDK API. Other libraries are private, and Android N only allows access to them for platform HALs, system daemons, and the like. If you aren’t sure whether your app uses private libraries, you can immediately check it for warnings on the N Developer Preview.
https://android-developers.googleblog.com/2016/06/improving-...
These stable APIs,
https://developer.android.com/ndk/guides/stable_apis
Also some obligatory Linux vs GNU/Linux comment. (and it's not like GNU/Linux doesn't ever change under your feet - see the glibc DT_HASH debacle)
Google relies on Linux LTS kernels. When the Linux LTS team dropped support from 6 years down to 2 years, Google stepped in to cover the 4-year gap.
It is Linux. It's basically a distro.
Someone can statically build a freestanding executable/so targetting arm64 linux (specifically the right android linux kernel version) and it will run fine on Android. The syscall interface, process model, file descriptors, signals, memory mapping, all of this is Linux, this is what people mean when they say Android is just Linux.
2026-04-29: Submitted detailed information about the rxrpc vulnerability and a weaponized exploit that achieves root privileges on Ubuntu to [email protected].
2026-04-29: Submitted the patch for the rxrpc vulnerability to the netdev mailing list. Information about this issue was published publicly.
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 the esp vulnerability were published publicly by an unrelated third party, breaking the embargo.
2026-05-07: After obtaining agreement from distribution maintainers to fully disclose Dirty Frag, the entire Dirty Frag document was published.
If there's no PoC, how can you really be sure?
But this is very similar to Copy Fail, and I'm assuming there was an assumption that others might also discover this soon as well. Hence the urgency.
At least that's my charitable interpretation.
However, it can be used to modify files that are passed into the container (e.g. Docker run -v), or files that are shared with other containers (e.g. other Docker containers sharing the same layers). kube-proxy with Kubernetes happens to share a trusted binary with containers by default, which is how it can be exploited: https://github.com/Percivalll/Copy-Fail-CVE-2026-31431-Kuber...
AI is neat because it's higher signal but yeah no, we're not getting anywhere close to "safe linux", AI or not.
An LPE only allows an attacker who can already execute code on the system to become root. So, bad, yes, but it doesn't mean you are immediately pwned.
you think the reporters and the distribution maintainers colluded to... get 5 minutes of attention?
that would be exceptionally stupid of the distribution maintainers and destroy all trust.
2. Bsds don’t have the same optimizations that Linux has. Bsds generally try to pursue corrrectness
That being said there were just a bunch of vulnerabilities in freebsd
macOS has had its own dirty cow attack and I know there’s for sure more memory ones just based on the way the xnu kernel works.
So no Linux isn’t really worse per say
- more people are using it (assuming macos is in its own bucket perhaps) - bigger surface areas (esp NetBSD has in my limited understanding just less stuff that can go boom) - more churn, ie more new stuff than can be buggy released more often.
Of course, because of that, more eyes are on Linux, so I'm not sure where that security tradeoff is.
Not criticizing whoever found the bug, of course.
That said, running every process in its own micro VM is looking more attractive by the minute.
But yes, micro VMs are a great idea!