Upload2Shell (CVE-2026-125396)

(pwn.ai)

131 points | by f311a 18 hours ago

19 comments

  • hannob 16 hours ago
    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.

    • viraptor 16 hours ago
      > 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

      • fock 7 hours ago
        requirements: Cairo, psutils, Pillow.
        • viraptor 6 hours ago
          Why is it interesting?
          • __del__ 5 hours ago
            i imagine those dependencies are non-trivial outside of linux
    • trebligdivad 15 hours ago
      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.
      • werrett 4 hours ago
        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

    • Natsu 2 hours ago
      I feel like the AI vulnerability flood is going to get called "2shell with that" after all the 4shell names.
    • koolba 13 hours ago
      > 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.

      • wolfi1 7 hours ago
        could be, but it's just ghostscript, rewritten
    • g-b-r 14 hours ago
      Hopefully few people install Ghostscript if they don't need it, on a server that processes user uploads
      • WesolyKubeczek 13 hours ago
        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.

        • viraptor 8 hours ago
          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.
          • asimovDev 6 hours ago
            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
            • g-b-r 4 hours ago
              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.

              • WesolyKubeczek 1 hour ago
                „When you can” is doing the heavy lifting here. Also, sometimes you need to process PostScript-adjacent things.
        • g-b-r 8 hours ago
          Yeah, sadly true
  • dabinat 18 hours ago
    This seems like important information but this page is completely unreadable.
    • TacticalCoder 17 hours ago
    • g-b-r 14 hours ago
      It's of course hidden behind a tiny link, but there's information at https://pwn.ai/upload2shell/research.html
    • CER10TY 15 hours ago
      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.

      • cyphar 8 hours ago
        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).

        • Dreami 5 hours ago
          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
        • Pokepokalypse 7 hours ago
          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.

  • int0x29 13 hours ago
    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.

    • Thorentis 12 hours ago
      I call these "slop disclosures". I have slop fatigue.
      • SV_BubbleTime 12 hours ago
        The first hint in <x>2Shell. The first twenty times it was ok, now, it’s just lazy.
        • DangIsATyrant 7 hours ago
          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)
        • J-Kuhn 2 hours ago
          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.

        • Bootvis 5 hours ago
          Indeed, why not Ghost in the Shell or Shell in the Ghost
          • necovek 5 hours ago
            Ghost Buster? As in, busting the ghostscript open.
  • megous 13 hours ago
    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.
    • milkshakes 8 hours ago
      > 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!

    • pastel8739 11 hours ago
      We should just use wasm for all this stuff
      • invalidname 7 hours ago
        Why not a restricted container?
  • dvh 14 hours ago
    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?
    • manytimesaway 12 hours ago
      For sure, having thousands of thousands of minimal libraries is truly what made NPM...
  • topranks 13 hours ago
    Seems like they did the right thing disclosure wise anyway.

    Who cares about the name all bugs get stupid names.

  • rietta 9 hours ago
    ghostscript is everywhere. It's also AGPL, which I am not certain is commonly understood.
  • freecodeio 14 hours ago
    this title isn't helping and is just clickbait
    • usef- 10 hours ago
      yes, and it feels like a mild abuse of the "2shell" name. It's a flaw in ghostscript, not the upload control.
  • dgacmu 16 hours ago
    Wish they'd named it Shell In The Ghost
    • skeledrew 15 hours ago
      Missed opportunity.
    • steve1977 16 hours ago
      We don't do human creativity around here sir.
    • GoblinSlayer 3 hours ago
      Just tell AI to do it.
  • layer8 15 hours ago
    I’m waiting for Stairway2Shell.
    • mpvjs 14 hours ago
      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.

    • xorcist 2 hours ago
      It's better than the inevitable highway2shell.
  • derac 17 hours ago
    this name is god awful. ask an llm to name your thing or something
  • varenc 10 hours ago
    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.
  • philipwhiuk 17 hours ago
    Weird title. Feels like it should be GhostScript2Shell.

    And the disclosure process is problematic if it's all the same issue.

    • fock 16 hours ago
      probably also the thumbnailer in your favourite linux-desktop. And the file-indexer ...
  • baggy_trough 17 hours ago
    If you want to use Ghostscript in a secure manner, you're going to have a bad time.
    • mpvjs 13 hours ago
      - Claude, examine and fix bugs in Ghostscript. - Thinking ........... ....................... ............................................... ..........................

      And this is how the AI rebellion started

      • WesolyKubeczek 13 hours ago
        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.
  • Xinti 13 hours ago
    [dead]
  • Xinti 13 hours ago
    [flagged]
    • tomhow 11 hours ago
      We've banned this account for going on tilt with a series of egregiously abusive comments over the past two days.
    • oguz-ismail2 11 hours ago
      imagine being this mad at 5 AM in the morning lol