Systemd lack of seat verification in pam module permits spoofing active session to polkit Vulnerability / Exploit
/
/
/
Exploits / Vulnerability Discovered : 2019-04-23 |
Type : dos |
Platform : linux
This exploit / vulnerability Systemd lack of seat verification in pam module permits spoofing active session to polkit is for educational purposes only and if it is used you will do on your own risk!
[+] Code ...
As documented at
<https://www.freedesktop.org/software/polkit/docs/latest/polkit.8.html>, for
any action, a polkit policy can specify separate levels of required
authentication based on whether a client is:
- in an active session on a local console
- in an inactive session on a local console
- or neither
This is expressed in the policy using the elements "allow_any",
"allow_inactive" and "allow_active". Very roughly speaking, the idea here is
to give special privileges to processes owned by users that are sitting
physically in front of the machine (or at least, a keyboard and a screen that
are connected to a machine), and restrict processes that e.g. belong to users
that are ssh'ing into a machine.
For example, the ability to refresh the system's package index is restricted
this way using a policy in
/usr/share/polkit-1/actions/org.freedesktop.packagekit.policy:
<action id="org.freedesktop.packagekit.system-sources-refresh">
[...]
<description>Refresh system repositories</description>
[...]
<message>Authentication is required to refresh the system repositories</message>
[...]
<defaults>
<allow_any>auth_admin</allow_any>
<allow_inactive>auth_admin</allow_inactive>
<allow_active>yes</allow_active>
</defaults>
</action>
On systems that use systemd-logind, polkit determines whether a session is
associated with a local console by checking whether systemd-logind is tracking
the session as being associated with a "seat". This happens through
polkit_backend_session_monitor_is_session_local() in
polkitbackendsessionmonitor-systemd.c, which calls sd_session_get_seat().
The check whether a session is active works similarly.
systemd-logind is informed about the creation of new sessions by the PAM
module pam_systemd through a systemd message bus call from
pam_sm_open_session() to method_create_session(). The RPC method trusts the
information supplied to it, apart from some consistency checks; that is not
directly a problem, since this RPC method can only be invoked by root.
This means that the PAM module needs to ensure that it doesn't pass incorrect
data to systemd-logind.
Looking at the code in the PAM module, however, you can see that the seat name
of the session and the virtual terminal number come from environment
variables:
This is actually documented at
<https://www.freedesktop.org/software/systemd/man/pam_systemd.html#Environment>.
After some fixup logic that is irrelevant here, this data is then passed to
the RPC method.
One quirk of this issue is that a new session is only created if the calling
process is not already part of a session (based on the cgroups it is in,
parsed from procfs). This means that an attacker can't simply ssh into a
machine, set some environment variables, and then invoke a setuid binary that
uses PAM (such as "su") because ssh already triggers creation of a session via
PAM. But as it turns out, the systemd PAM module is only invoked for
interactive sessions:
# cat /usr/share/pam-configs/systemd
Name: Register user sessions in the systemd control group hierarchy
Default: yes
Priority: 0
Session-Interactive-Only: yes
Session-Type: Additional
Session:
optional pam_systemd.so
So, under the following assumptions:
- we can run commands on the remote machine, e.g. via SSH
- our account can be used with "su" (it has a password and isn't disabled)
- the machine has no X server running and is currently displaying tty1, with
a login prompt
we can have our actions checked against the "allow_active" policies instead of
the "allow_any" policies as follows:
- SSH into the machine
- use "at" to schedule a job in one minute that does the following:
* wipe the environment
* set XDG_SEAT=seat0 and XDG_VTNR=1
* use "expect" to run "su -c {...} {our_username}" and enter our user's
password
* in the shell invoked by "su", perform the action we want to run under the
"allow_active" policy
I tested this in a Debian 10 VM, as follows ("{{{...}}}" have been replaced),
after ensuring that no sessions are active and the VM's screen is showing the
login prompt on tty1; all following commands are executed over SSH:
1 sessions listed.
normal_user@deb10:~$ pkcon refresh -p </dev/null
Transaction: Refreshing cache
Status: Waiting in queue
Status: Waiting for authentication
Status: Finished
Results:
Fatal error: Failed to obtain authentication.
normal_user@deb10:~$ at -f /home/normal_user/session_outer.sh {{{TIME}}}
warning: commands will be executed using /bin/sh
job 25 at {{{TIME}}}
{{{ wait here until specified time has been reached, plus time for the job to finish running}}}
normal_user@deb10:~$ cat /tmp/atjob.log
===== ENV DUMP =====
XDG_SEAT=seat0
XDG_VTNR=1
PWD=/home/normal_user
===== SESSION_OUTER =====
10:memory:/system.slice/atd.service
9:freezer:/
8:pids:/system.slice/atd.service
7:perf_event:/
6:devices:/system.slice/atd.service
5:net_cls,net_prio:/
4:cpuset:/
3:blkio:/
2:cpu,cpuacct:/
1:name=systemd:/system.slice/atd.service
0::/system.slice/atd.service
===== OUTER LOGIN STATE =====
SESSION UID USER SEAT TTY
7 1001 normal_user pts/0
1 sessions listed.
===== MIDDLE TESTING PKCON
Transaction: Refreshing cache
Status: Waiting in queue
Status: Waiting for authentication
Status: Finished
Results:
Fatal error: Failed to obtain authentication.
===== INNER LOGIN STATE =====
SESSION UID USER SEAT TTY
18 1001 normal_user seat0 pts/1
7 1001 normal_user pts/0