Comment by hughw

14 hours ago

Ran as a fresh new default user in a ubuntu:latest container

  git clone https://github.com/V4bel/dirtyfrag.git && cd dirtyfrag && gcc -O0 -Wall -o exp exp.c -lutil && ./exp

Result:

  dirtyfrag: failed (rc=3)

Good news!

I got the same running it inside a container, but got a shell when running it directly in the host. This only shows that the exploit doesn't work inside a container. So, containers aren't vulnerable, or the script needs some adjustments to make it work in containers.

Since copy fail can be used to escape containers (https://github.com/Percivalll/Copy-Fail-CVE-2026-31431-Kuber...), I'm guessing the exploit needs some changes only.

  • The repo you linked works by replacing files that are being used by other privileged containers on the same system. That works for the Kubernetes case (I'm a little surprised they don't use static binaries for their own privileged containers, seems a little dangerous to share any kind of data with untrusted tenants even if it's read-only) but not standalone containers.

    However, there is a much an easier way of doing a breakout -- you can corrupt the host runc binary in a way analogous to CVE-2019-5736. The next time a container is spawned, the host runc binary will get run as as root and that's that.

    Ironically, the first version of the protection against this attack I wrote also protected against page cache poisoning (by making a temporary copy of the runc binary during container setup in a sealed memfd and re-execing that) but the runtime cost of copying a 10MB binary at container startup was seen as too expensive by some users[1] so we ended up with a setup that shares the same page cache. I also distinctly remember arguing at the time that something like Dirty Cow could always happen in the future, and the memfd approach was better for that reason -- maybe I should've stuck to my guns more... :/

    In practice the solution for containers is to update your seccomp policy to block the vulnerable syscall.

    [1]: https://github.com/opencontainers/runc/issues/1980

Wouldn't count on container being a reliable testing platform for this. Loads of stuff - legitimate or otherwise - fails in containers