It's been known for a while, and certainly long before the recent AI vuln flood, that ghostscript is really problematic in terms of security.
Ultimately, postscript is its own programming language, and executing a programming language in a secure, sandboxed way is not trivial. ghostscript has only a very poor pseudo sandbox that fails regularly. Unfortunately, ghostscript is the only implementation of postscript that's out there.
All of that would matter much less if people only used it to parse postscript. But it's used indirectly by imagemagick (you can disable it), which means many people using imagemagick to parse images from untrusted sources get exposed to ghostscript's poor security.
> Unfortunately, ghostscript is the only implementation of postscript that's out there.
That needs lots of qualifiers. It's the only popular and widely distributed open-source implementation with known good performance and wide compatibility.
There were so many other implementations and vibed ports already that I expect ghostscript to be actually replaced in distros within a year or two. Very random example https://github.com/AndyCappDev/postforge
Yes, and people didn't previously write big webpages, come up with a fancy name, and notify about a dozen different companies before just writing a ghostscript fix.
No previously it was just confined to OpenSSL bugs (2014) [1], linux kernel race conditions (2016) [2], CPU speculative execution (2018) [3] just to randomly list some of the earlier ones.
It quickly became cringe, but the annoyance you're feeling right now is not due to folks branding vulns.. It's that the rate of these bugs has gone up 5-20x since ~July this year.
Unfortunately this is the new normal. It's not going to get "better" any time soon.
You install ImageMagick, it most likely depends on Ghostscript.
You use something else than ImageMagick, say, to resize an uploaded image, and the moment it sees a color-managed image, it will produce something that looks like puke.
I had to do a very custom build of ImageMagick to get rid of dependencies on libgs and libX11. It wasn't fun.
ImageMagick was problematic for so long by pulling everything and the kitchen sink as dependencies. It's just not something you'd like to replace because it's got nice features and... works. It feels absurd but right now, if you have a very narrow use case (like just resizing), it may be better to make an LLM create an alternative, in a safer language, with validated core components, tested on a large number of samples.
Another alternative is GraphicsMagick (http://www.graphicsmagick.org/) which aims to be API and utility compatible with ImageMagick, but spends most of its development effort on fuzzing, verification, etc. It's a more secure alternative to ImageMagick for those who've developed something significant that depends on ImageMagick.
I was looking through the PPDs a Brother printer installed for Linux (trying to get it working on FreeBSD), it had a pre-built ghostscript binary in there that hadn't been updated in god knows how long. I wonder if printing a PDF with an embedded font or something could trigger that.
Seriously, it's just such a bad design. Extremely dense but at the same time extremely devoid of any info. Sorry for the broad negativity here, but this page just screams "I want CVE cred" without providing any useful info on whether this is a "patch now pants on fire" or "yeah maybe next cycle".
At least they didn't lead with a useless CVSS score like all the other vulnerabilities we see popping up now.
This is just what slop sites look like. The LLM doesn't "understand" that the reader doesn't have it's context window in front of them when reading their output (i.e., the webpage) because to it everything is in the same bucket. IMHO this is probably because the attention of the transformer model by design lights up details that are relevant to the model but irrelevant to the reader.
Same goes for basically any descriptive text they output (documentation, comments, commit messages, etc).
It makes me feel dumb when I'm trying to read this stuff. It actually took me a while to figure out what it was. I kind of thought I was going crazy. (esp since these characteristics would change from one version to another; particularly Opus 4.5 to 4.7, then to 5.0, and now 5.5 seems to be a step backwards.
You can tell it until you're blue in the face to be more explicit, but it's always re-active, and it just seems like you can't make it proactively be more explicit and clear.
Some days I have to really struggle to get through just one page of text from Claude trying to explain something to me. Especially when it's a complex problem with a lot of small components to it.
I also experience it in the other direction. When you build an app and it guesses a relation wrongly, then you tell it “this is wrong, it works like X, not like Y”. The LLM then cannot stop itself from writing that into every user-facing description where this relation comes up. As a user it then sounds, as if a car manual was written: to accelerate the car, press the throttle (not the steering wheel). I then have to spend prompts telling it to only describe the resulting spec, not the way we worked towards it, and that (quite high) concept then gets lost again the next time some description needs to be written
Yeah I see that a lot too and I think it's a symptom of the same underlying problem -- to the model, the entire context window is incredibly important and it doesn't "understand" that it's utterly irrelevant to almost anyone else (where "doesn't understand" actually means "this is a pathology of the attention architecture").
The best part is that if you ask it to spawn subagents to get cleaner contexts, it will also very often decide to pollute the subagent prompt with the same crap. And because the resulting text reads as unhinged (read: statistically unlikely) the subagent model overweights its importance.
That timeline is unreadable. Just say what was found impacted when with versions along with patch dates and versions. I don't need to know about case studies that will hypothetically be written. I don't need little blurbs under everything. "What we said" should instead be "publicly disclosed."
Honestly the whole thing should be a table so you can group the different bits of the timeline for different software/services by software/service rather than time.
Someone on twitter looked up every 2shell domain registered in the last month, there's still like 30 more of these in the pipe (or someone just collecting)
Not only unreadable, but it's also in very poor taste to use those kinds of exciting, product-announcement-like fonts and look-and-feels for a security issue. That kind of design is meant to convey the "isn't this fun, guys?" feeling.
ghostscript, imagemagick, ffmpeg, etc. etc. all these passthrough data in data out programs that don't need much of anything from the system, I'd think people would run them with untrusted input in a tight whitelist based seccomp+landlock sandbox. it's a bit annoying if they are not statically compiled, due to ld.so loading libraries, but possible.
Do you think people will finally start removing decade old cruft from libraries and/or start creating single purpose minimal libraries? Or will the defense be just reactionary and on per cve basis?
"Mishell" by The Byteless, about endianness (these are "words" that go together well)
"Ring my shell" by Anita Reward, about bounties for remote terminal exploits (I guess You did by the look in your eyes - Look in your eyes, look in your eyes)
"Screen is blue" a classic by Andre Popp Upp about the well known behavior of windows OS family.
I'm dubious of some of the claims in the timeline. I have no doubt they found an RCE vuln, but people know that your document rendering service should be sandboxed (beyond Ghostscript's sandbox). So it's a bit disingenuous not to acknowledge that their RCEs for companies like Google and Dropbox almost certainly only get them into tightly sandboxed servers that lacked any sensitive access, exactly for this reason.
Yes but no: Since GhostScript is used in a lot of other tooling it can affect those as well.
So if you're using imagemagick to resize user uploaded jpgs or pngs, an attacker might be able to reach GS by sending you a PostScript with the png extension.
So yes, technically it's all just GhostScript, but the impact is much wider than gs2shell would suggest.
(I personally don't care about names and brandings, but there are worse and better choices)
Theoretically, you could make a PostScript program that would output an image of a robot arm strangling its human handler, while also directing said robot arm to strangle its human handler.
It's been known for a while, and certainly long before the recent AI vuln flood, that ghostscript is really problematic in terms of security.
Ultimately, postscript is its own programming language, and executing a programming language in a secure, sandboxed way is not trivial. ghostscript has only a very poor pseudo sandbox that fails regularly. Unfortunately, ghostscript is the only implementation of postscript that's out there.
All of that would matter much less if people only used it to parse postscript. But it's used indirectly by imagemagick (you can disable it), which means many people using imagemagick to parse images from untrusted sources get exposed to ghostscript's poor security.
> Unfortunately, ghostscript is the only implementation of postscript that's out there.
That needs lots of qualifiers. It's the only popular and widely distributed open-source implementation with known good performance and wide compatibility.
There were so many other implementations and vibed ports already that I expect ghostscript to be actually replaced in distros within a year or two. Very random example https://github.com/AndyCappDev/postforge
requirements: Cairo, psutils, Pillow.
Why is it interesting?
i imagine those dependencies are non-trivial outside of linux
They're trivial anywhere you have Python.
they're also a trivial way to call all amounts of random C-code with... Probably even gs...
Once you want to support everything ghostscript does, nothing is trivial. Ghostscript depends on fun stuff too https://packages.ubuntu.com/stonking/libgs10
But it could be worse. Cairo and Pillow are very established.
Yes, and people didn't previously write big webpages, come up with a fancy name, and notify about a dozen different companies before just writing a ghostscript fix.
No previously it was just confined to OpenSSL bugs (2014) [1], linux kernel race conditions (2016) [2], CPU speculative execution (2018) [3] just to randomly list some of the earlier ones.
It quickly became cringe, but the annoyance you're feeling right now is not due to folks branding vulns.. It's that the rate of these bugs has gone up 5-20x since ~July this year.
Unfortunately this is the new normal. It's not going to get "better" any time soon.
[1]: https://www.heartbleed.com/ [2]: https://dirtycow.ninja/ [3]: https://spectreattack.com
Oh I still regard 2014 as modern; remember Ghostscript came out in 1988 - I've been using it since the mid 90's!
Hopefully few people install Ghostscript if they don't need it, on a server that processes user uploads
You install ImageMagick, it most likely depends on Ghostscript.
You use something else than ImageMagick, say, to resize an uploaded image, and the moment it sees a color-managed image, it will produce something that looks like puke.
I had to do a very custom build of ImageMagick to get rid of dependencies on libgs and libX11. It wasn't fun.
Yeah, sadly true
ImageMagick was problematic for so long by pulling everything and the kitchen sink as dependencies. It's just not something you'd like to replace because it's got nice features and... works. It feels absurd but right now, if you have a very narrow use case (like just resizing), it may be better to make an LLM create an alternative, in a safer language, with validated core components, tested on a large number of samples.
Something that uses CImg just for resizing could be a decent alternative. It’s a library but they provided a nice cli for it the last time i checked
Yes, it seems much smaller than ImageMagick.
ImageMagick has a big attack surface, when most people probably use it just for cropping and resizing.
Just use something leaner, when you can.
„When you can” is doing the heavy lifting here. Also, sometimes you need to process PostScript-adjacent things.
Another alternative is GraphicsMagick (http://www.graphicsmagick.org/) which aims to be API and utility compatible with ImageMagick, but spends most of its development effort on fuzzing, verification, etc. It's a more secure alternative to ImageMagick for those who've developed something significant that depends on ImageMagick.
I was looking through the PPDs a Brother printer installed for Linux (trying to get it working on FreeBSD), it had a pre-built ghostscript binary in there that hadn't been updated in god knows how long. I wonder if printing a PDF with an embedded font or something could trigger that.
> Unfortunately, ghostscript is the only implementation of postscript that's out there.
This sounds like an afternoon’s worth of prompting could double the number of implementations.
could be, but it's just ghostscript, rewritten
I feel like the AI vulnerability flood is going to get called "2shell with that" after all the 4shell names.
A lot of these 4shell vulns are only theoretical RCE. That's not to say they aren't serious, but I think the severity is exaggerated a fair amount.
GS should be pledged under OpenBSD.
This seems like important information but this page is completely unreadable.
For those on Debian, the usual:
https://security-tracker.debian.org/tracker/CVE-2026-125396
Seriously, it's just such a bad design. Extremely dense but at the same time extremely devoid of any info. Sorry for the broad negativity here, but this page just screams "I want CVE cred" without providing any useful info on whether this is a "patch now pants on fire" or "yeah maybe next cycle".
At least they didn't lead with a useless CVSS score like all the other vulnerabilities we see popping up now.
This is just what slop sites look like. The LLM doesn't "understand" that the reader doesn't have it's context window in front of them when reading their output (i.e., the webpage) because to it everything is in the same bucket. IMHO this is probably because the attention of the transformer model by design lights up details that are relevant to the model but irrelevant to the reader.
Same goes for basically any descriptive text they output (documentation, comments, commit messages, etc).
It makes me feel dumb when I'm trying to read this stuff. It actually took me a while to figure out what it was. I kind of thought I was going crazy. (esp since these characteristics would change from one version to another; particularly Opus 4.5 to 4.7, then to 5.0, and now 5.5 seems to be a step backwards.
You can tell it until you're blue in the face to be more explicit, but it's always re-active, and it just seems like you can't make it proactively be more explicit and clear.
Some days I have to really struggle to get through just one page of text from Claude trying to explain something to me. Especially when it's a complex problem with a lot of small components to it.
I also experience it in the other direction. When you build an app and it guesses a relation wrongly, then you tell it “this is wrong, it works like X, not like Y”. The LLM then cannot stop itself from writing that into every user-facing description where this relation comes up. As a user it then sounds, as if a car manual was written: to accelerate the car, press the throttle (not the steering wheel). I then have to spend prompts telling it to only describe the resulting spec, not the way we worked towards it, and that (quite high) concept then gets lost again the next time some description needs to be written
Yeah I see that a lot too and I think it's a symptom of the same underlying problem -- to the model, the entire context window is incredibly important and it doesn't "understand" that it's utterly irrelevant to almost anyone else (where "doesn't understand" actually means "this is a pathology of the attention architecture").
The best part is that if you ask it to spawn subagents to get cleaner contexts, it will also very often decide to pollute the subagent prompt with the same crap. And because the resulting text reads as unhinged (read: statistically unlikely) the subagent model overweights its importance.
It's of course hidden behind a tiny link, but there's information at https://pwn.ai/upload2shell/research.html
That timeline is unreadable. Just say what was found impacted when with versions along with patch dates and versions. I don't need to know about case studies that will hypothetically be written. I don't need little blurbs under everything. "What we said" should instead be "publicly disclosed."
Honestly the whole thing should be a table so you can group the different bits of the timeline for different software/services by software/service rather than time.
I call these "slop disclosures". I have slop fatigue.
The first hint in <x>2Shell. The first twenty times it was ok, now, it’s just lazy.
Someone on twitter looked up every 2shell domain registered in the last month, there's still like 30 more of these in the pipe (or someone just collecting)
Indeed, why not Ghost in the Shell or Shell in the Ghost
Ghost Buster? As in, busting the ghostscript open.
The first one was in log4j, so that sounded similar to log4shell.
But now, even things that are not 4j get the 4shell "pun". But... it doesn't work.
Not only unreadable, but it's also in very poor taste to use those kinds of exciting, product-announcement-like fonts and look-and-feels for a security issue. That kind of design is meant to convey the "isn't this fun, guys?" feeling.
ghostscript, imagemagick, ffmpeg, etc. etc. all these passthrough data in data out programs that don't need much of anything from the system, I'd think people would run them with untrusted input in a tight whitelist based seccomp+landlock sandbox. it's a bit annoying if they are not statically compiled, due to ld.so loading libraries, but possible.
We should just use wasm for all this stuff
Why not a restricted container?
Too much surface area for bugs
Maybe ok for low volume stuff but you think YouTube is going to take a 2x performance hit for all their video processing?
Lightweight VMs are probably the best solution in that case.
> I'd think people would run them with untrusted input in a tight whitelist based seccomp+landlock sandbox.
in fact, many do. if you're not doing this all your unsafe file processing in 2026, you're in for a bad time!
OpenBSD will get tons of support with pledged graphicsmagick and ffmpeg (and maybe GS).
I wish Ori B finishes the GeFS port from 9front.
Do you think people will finally start removing decade old cruft from libraries and/or start creating single purpose minimal libraries? Or will the defense be just reactionary and on per cve basis?
For sure, having thousands of thousands of minimal libraries is truly what made NPM...
like left-pad :PP
A GhostScript bug (from a very outdated version) famously led to the compromise of 4chan.
Seems like they did the right thing disclosure wise anyway.
Who cares about the name all bugs get stupid names.
this title isn't helping and is just clickbait
yes, and it feels like a mild abuse of the "2shell" name. It's a flaw in ghostscript, not the upload control.
Wish they'd named it Shell In The Ghost
We don't do human creativity around here sir.
Missed opportunity.
Just tell AI to do it.
No, only hackneyed titles allowed.
I’m waiting for Stairway2Shell.
In the meantime you may enjoy the playlist:
"Mishell" by The Byteless, about endianness (these are "words" that go together well)
"Ring my shell" by Anita Reward, about bounties for remote terminal exploits (I guess You did by the look in your eyes - Look in your eyes, look in your eyes)
"Screen is blue" a classic by Andre Popp Upp about the well known behavior of windows OS family.
It's better than the inevitable highway2shell.
this name is god awful. ask an llm to name your thing or something
I'm dubious of some of the claims in the timeline. I have no doubt they found an RCE vuln, but people know that your document rendering service should be sandboxed (beyond Ghostscript's sandbox). So it's a bit disingenuous not to acknowledge that their RCEs for companies like Google and Dropbox almost certainly only get them into tightly sandboxed servers that lacked any sensitive access, exactly for this reason.
Weird title. Feels like it should be GhostScript2Shell.
And the disclosure process is problematic if it's all the same issue.
probably also the thumbnailer in your favourite linux-desktop. And the file-indexer ...
Yes but no: Since GhostScript is used in a lot of other tooling it can affect those as well.
So if you're using imagemagick to resize user uploaded jpgs or pngs, an attacker might be able to reach GS by sending you a PostScript with the png extension.
So yes, technically it's all just GhostScript, but the impact is much wider than gs2shell would suggest.
(I personally don't care about names and brandings, but there are worse and better choices)
ghostscript is everywhere. It's also AGPL, which I am not certain is commonly understood.
If you want to use Ghostscript in a secure manner, you're going to have a bad time.
- Claude, examine and fix bugs in Ghostscript. - Thinking ........... ....................... ............................................... ..........................
And this is how the AI rebellion started
Theoretically, you could make a PostScript program that would output an image of a robot arm strangling its human handler, while also directing said robot arm to strangle its human handler.