DEV Community

Cover image for Isolation Is Not Virtualization: Auditing Docker Sockets and Layer Leaks
Mohamed
Mohamed

Posted on

Isolation Is Not Virtualization: Auditing Docker Sockets and Layer Leaks

Containers give you isolation, not virtualization. Confuse the two and a web-container command injection becomes full host root.

Containers do not provide virtualization. They provide isolation. Treat isolation as a hard security boundary and a low-privilege injection inside a web container can become root on the host.

One of the most battle-tested tools for finding these misconfigs is dockerscan by cr0hn:

https://github.com/cr0hn/dockerscan

Three misconfigurations cause most container breaches.

1. The exposed Docker socket

Teams mount the host socket into a container as a Docker-in-Docker shortcut:

# catastrophic:
volumes:
  - /var/run/docker.sock:/var/run/docker.sock
Enter fullscreen mode Exit fullscreen mode

Write access to the socket is root on the host — no kernel exploit needed:

docker run -v /:/host -it alpine chroot /host
Enter fullscreen mode Exit fullscreen mode

That mounts the entire host filesystem and drops you into a root shell.

2. Privileged containers

The --privileged flag strips almost all isolation: AppArmor/SELinux off, all of /dev exposed, dangerous capabilities (CAP_SYS_ADMIN, CAP_SYS_MODULE) granted. From there an attacker can mount disks, load kernel modules, or escape via the cgroup-v1 release_agent trick.

3. Image-layer secret leakage

Deleting a secret in a later step does not remove it from the image:

RUN export AWS_SECRET_KEY="AKIA..." && ./build.sh
RUN rm -rf /secrets.env   # only writes a whiteout layer
Enter fullscreen mode Exit fullscreen mode

Image layers are immutable. Tools like dockerscan read every historical layer and pull out tokens, SSH keys and passwords in seconds.

The hardened checklist

  • Never mount the Docker socket. Use rootless builds, Kaniko, or isolated runner VMs.
  • Drop all capabilities (--cap-drop=ALL) and add back only what the binary needs (CAP_NET_BIND_SERVICE).
  • Run as non-root — USER appuser, never UID 0.
  • Keep secrets out of layers with multi-stage builds or BuildKit --mount=type=secret.

When did you last audit your production images and daemon config? Comment below.

Top comments (0)