• pcouy@lemmy.pierre-couy.frOP
    link
    fedilink
    English
    arrow-up
    1
    ·
    2 hours ago

    Yeah, clearly the “2root” part has nothing to do with nginx. However, given how quickly linux privescs are dropping these days, I think that even on fully patched systems it’s only a matter of time before any shell turns into root. Also, upgrading the kernel requires a reboot, and many server operators are quite reluctant to reboot there boxes in fear of breaking something. Plenty of servers are running an outdated kernel

    I’ve stopped considering Linux users or containers as any kind of security boundary a few months ago

    • redjard@reddthat.com
      link
      fedilink
      English
      arrow-up
      2
      ·
      2 hours ago

      Even calling this “nginx2shell” would be very clickbaity. This isn’t a realistically usable attack. Only one exploit can ever be a “2shell” for any software, it should be a once in a decade vulnerability getting that name. Even if it was escalatable, and even if it didn’t require extensive resources for every exploitation, it’d steíll depend on only semi-common configuration. Nothing deserving of “2shell” on multiple levels.

      With the privesc, I would look into that again. I got the impression we were through the major escalation pathways, and now have increasingly convoluted and less plausible leftovers being reported.
      Even if they report 1k garbage exploits at once, that’s still not gonna do much for any actual real-world risk to normal hosters.
      Given the marketing they do when something is even theoretically plausible as an escallation, and given the worst I recall hearing recently was one of the tcp ones that only “worked” for sandbox, container, and vm escape (and ofc user to root), I don’t think we have that many more left. Maybe ai companies are holding a few back for another big press release or two, we’ll see. If this bundled root exploit turns out to be a major dud, I doubt they have much material left.

      AI companies have an interest to embed the general sentiment that ai exploitation is growing exponentially, and the only way forward is to buy ai and keep testing and developing software with it to protect from ai attacks ahead of time.
      But in practice in the existing fields ai was introduced to, there was a saturation. Initial gains when making ai capable of working in the field and it then jumping from a few categories to most inside that field. and then some more for more specific training for the tasks, and then not much past that.

      We had a few classes of bugs AI was suddenly very capable at finding. We probably got all the faulty page handling in spring, then various use after free and general memory handling bugs. We’ll probably get ai implementing existing testing suites more widely across software. Like how the 4-cert was found by just running the normal testing tool on some random cert companies obscure intermediate certs, or how some basic sanity checks found the musl ssh password thing.
      Once memory saftey has been improved to the level current llms can understand, and basic security tests have been written for software, we might extend this to continuously more obscure software but eventually there’d not be much more to report.

      My expectation is AI companies are stockpiling the stuff they have, releasing it on a schedule and with marketing that suggest an ever worsening problem, and then hope enough people buy ai products that they can then blame the sudden reduction on ai found problems on the new ai testing.

      To be fair, this is a bit conspiritorial, but even taking the exact opposite view, there’s still only so many bugs to find. If you think ai will keep growing in bug finding ability, then at some point they’ll find all, and the open weight models will also find all, and then eventually this all becomes a unit test that runs before stuff gets pushed.
      In my own expected scenario where llms are good at basic memory safety and logic bugs, that’s equally be the end result. Using some specialized local model as a compile-time check stand-in for something like the rust borrow checker or lean logic checker for languages that suffer from shit memory protections (looking at C++) or encourage bad logic patterns (looking at basically anything other than proof checkers and maybe rust).