The polkit unix-process subject for authorization is racy. It depended
on the (PID, startup_time) pair to be passed to polkit which then used /proc/PID/status
to find out the UID the process belongs to. Meanwhile the process could
have started a suid or pkexec process, changing the euid and/or uid at will.
The startup_time does not protect here, as its not changed across an execve().
Using /proc/PID/loginuid wont work either, as one could abuse fork-spawning
processes such as sshd, apache etc. to re-use recently freed process slots,
faking the loginuid. startup_time would theoretically help here, yet as
its not atomically passed along the message which is subject to polkit
authorization, the privileged process needs to learn it by looking up
/proc/PID/, which is racy again.
Therefore the only thing that could be used is the UID that is passed
atomically in the peer cred struct when receiving the message in question.
The whole thing needs fixing in polkit, to deprecate PID authorization
as well as several core packages to make use of the new API, or use
After discussing with upstream, Colin Walters made this private git of patches
Feel free to suggest improvements if necessary.
As required by list policy, I request a CRD of Sept 11th.
We also need CVE's assigned.
A PoC with example client/server which demonstrates the race
can be found here (it basically simulates libvirtd's way of
This is a (probably preliminary) list of CVEs and applications affected:
CVE-2013-4288 polkit: unix-process subject for authorization is racy
CVE-2013-4311 libvirt: insecure calling of polkit via libgobject API
CVE-2013-4324 spice-gtk: use of insecure polkit libgobject-1 API
CVE-2013-4325 hplip: use of insecure polkit DBUS API
CVE-2013-4326 rtkit: use of insecure polkit DBUS API
CVE-2013-4327 systemd: use of insecure polkit DBUS API
I suppose it's this?
Plus Cardoe committed 0.112 to tree today for this CVE.
I know at least rtkit, hplip, and systemd are all vulnerable to this.