http://blog.docker.com/2014/06/docker-container-breakout-pro...But nothing is stopping you from creating a binary and running it as root that perhaps mounts those filesystems, or sends ioctls / badness to the kernel.
I'm just shooting from the hip here and might be wrong, but I'm almost certain devices aren't namespaced. If you were to mknod a block device with the right major / minor numbers as the host's sda/sdb/etc, what would stop you from mounting the host's root filesystem and making changes?
Granted the "work around" is to drop the privileges of the containers as much as possible, use sVirt (via SELinux), drop as many capabilities as possible, don't run untrusted containers, etc. But this isn't a theoretical security problem, it is a real big attack surface.
Edit:
http://www.projectatomic.io/blog/2014/09/yet-another-reason-...