mount destinations can be swapped via symlink-exchange to cause mounts outside the rootfs

Summary

runc 1.0.0-rc94 and earlier are vulnerable to a symlink exchange attack whereby an attacker can request a seemingly-innocuous container configuration that actually results in the host filesystem being bind-mounted into the container (allowing for a container escape). CVE-2021-30465 has been assigned for this issue.

An attacker must have the ability to start containers using some kind of custom volume configuration, and while recommended container hardening mechanisms such as LSMs (AppArmor/SELinux) and user namespaces will restrict the amount of damage an attacker could do, they do not block this attack outright. We have a reproducer using Kubernetes (and the below description mentions Kubernetes-specific paths), but this is not a Kubernetes-specific issue.

The now-released runc v1.0.0-rc95 contains a fix for this issue, we recommend users update as soon as possible.

Details

In circumstances where a container is being started, and runc is mounting inside a volume shared with another container (which is conducting a symlink-exchange attack), runc can be tricked into mounting outside of the container rootfs by swapping the target of a mount with a symlink due to a time-of-check-to-time-of-use (TOCTTOU) flaw. This is fairly similar in style to previous TOCTTOU attacks (and is a problem we are working on solving with libpathrs).

However, this alone is not useful because this happens inside a mount namespace with MS_SLAVE propagation applied to / (meaning that the mount doesn't appear on the host -- it's only a "host-side mount" inside the container's namespace). To exploit this, you must have additional mount entries in the configuration that use some subpath of the mounted-over host path as a source for a subsequent mount.

However, it turns out with some container orchestrators (such as Kubernetes -- though it is very likely that other downstream users of runc could have similar behaviour be accessible to untrusted users), the existence of additional volume management infrastructure allows this attack to be applied to gain access to the host filesystem without requiring the attacker to have completely arbitrary control over container configuration.

In the case of Kubernetes, this is exploitable by creating a symlink in a volume to the top-level (well-known) directory where volumes are sourced from (for instance, /var/lib/kubelet/pods/$MY_POD_UID/volumes/kubernetes.io~empty-dir), and then using that symlink as the target of a mount. The source of the mount is an attacker controlled directory, and thus the source directory from which subsequent mounts will occur is an attacker-controlled directory. Thus the attacker can first place a symlink to / in their malicious source directory with the name of a volume, and a subsequent mount in the container will bind-mount into the container.