This issue affects OpenSSH if privilege separation is disabled (config option
UsePrivilegeSeparation=no). While privilege separation is enabled by default, it
is documented as a hardening option, and therefore disabling it should not
directly make a system vulnerable.
OpenSSH can forward TCP sockets and UNIX domain sockets. If privilege separation
is disabled, then on the server side, the forwarding is handled by a child of
sshd that has root privileges. For TCP server sockets, sshd explicitly checks
whether an attempt is made to bind to a low port (below IPPORT_RESERVED) and, if
so, requires the client to authenticate as root. However, for UNIX domain
sockets, no such security measures are implemented.
This means that, using "ssh -L", an attacker who is permitted to log in as a
normal user over SSH can effectively connect to non-abstract unix domain sockets
with root privileges. On systems that run systemd, this can for example be
exploited by asking systemd to add an LD_PRELOAD environment variable for all
following daemon launches and then asking it to restart cron or so. The attached
exploit demonstrates this - if it is executed on a system with systemd where
the user is allowed to ssh to his own account and where privsep is disabled, it
yields a root shell.
Proof of Concept: