Exploring Docker Security

$ whoami
● Patrick Kleindienst
● Computer Science & Media (CS3)
● student trainee at Bertsch Innovation GmbH (since 2014)
● interested in Linux, software development, infrastructure etc.
2

Outline
● Review: Hardware virtualization and VMs
● Docker at a glance
● Container internals (using the example of Docker)
● Container security: How secure is Docker?
● Conclusion and further thoughts
● Discussion
3

Review:Hardware
virtualization and
VMs
● Virtual Machine basics
● Pros and Cons
4

Review: Virtual Machine basics
5
● VM = replication of a computer system
● runs a whole operating system with its
own OS kernel
● hypervisor creates a virtual environment
for each VM (RAM, CPU, Storage, ..)
● hypervisor as an abstraction layer
between host and guest(s)
● each host may run multiple guest VMs

Hardware Virtualization: Pros and Cons
single kernel per VM offers high degree of
isolation
hypervisor reduces attack surface
VM escape is considered very difficult
improvement of hardware resources
utilization
guest OS may be different from host OS
● full kernel = almost certainly many bugs
● hypervisor may also ship with bugs
● not as efficient as an ordinary host
● running on virtual hardware is slower than
physical hardware
● highly elastic infrastructure based on VMs
is not so easy6

Docker at a
glance
● About Docker
● The Container approach
● Docker architecture
● Demo
7

About Docker
● started as dotCloud (shipping Software with LXC)
● release of Docker as Open Source Project (2013)
● slogan: “Build, Ship, Run”
● ease of packaging and deploying applications
● focused on usability
● trigger for DevOps movement
8

The Container approach
9
● no more hypervisor, no more VMs
● lightweight Docker Engine running on top
of host OS
● Docker engine runs apps along with their
dependencies as isolated processes
sharing the host kernel (Containers)
● Starting/Stopping a container takes
seconds instead of minutes (or even
hours)

Docker architecture (1)
Docker Image:
read-only template containing a minimal OS
(e.g. Ubuntu, Debian, CentOS, ..)
may also contain additional layers (JRE,
Python, Apache, VIM, ..)
published and shared by means of Dockerfiles
Docker Container:
additional read-write layer on top of an image
does not manipulate the underlying image
10

Docker architecture (2)
● Docker Client:
○ for interaction with Docker Daemon
○ shares a UNIX socket with the daemon
● Docker Daemon:
○ connects to the same UNIX socket as the
client
○ responsible for starting, stopping and
monitoring containers
11

What we’ve learned so far:
● In contrast to VMs, containers running on the same host share the
underlying kernel
● Therefore, they’re lightweight and save lots of resources
● As for starting/stopping/setup, they’re also much faster than traditional
VMs
● Docker distinguishes between Images and Containers
● Docker Images ship with at least a single minimal OS layer12

What we DON’T know so far:
Thinking about the underlying technology:
What exactly are file system layers and Copy-on-Write?
How to provide isolation between multiple containers running on same host?
Did Docker really invent all this stuff??
Thinking about security:
Eeehm, … how secure is running a container in the first place?
And what about Docker?13

Container internals
(using the
example of
Docker)
● Union file systems
● Namespaces
● Control groups
● Putting it all together
14

Union file systems (AUFS)
● unification filesystem: stack of multiple directories on an Linux host which
provides a single unified view (like stacked sheets on a overhead
projector)
● involved directories need a shared mount point (union mount)
● shared mount point provides a single view on the mounted directories
● a directory participating in a union mount is called a branch
● result: each layer simply stores what has changed compared to the layers15

Namespaces
● isolation mechanism of the Linux kernel
● provide processes with a different views on global resources
● examples: PIDs, network interfaces, mount points
● processes can work on that views without affecting the global
configuration
● Linux makes use of certain system calls for namespace creation
18

Mount Namespaces
19
● Linux OS maintains data structure
containing all existing mount points
(which fs is mounted on which path?)
● Kernel allows for cloning this data
structure and pass is to a process or
group of processes
● Process(es) can change their mount
points without side-effects
● e.g. allows for changing the root fs
(similar to chroot)

PID Namespaces
20
● in a single process tree, a privileged process
may inspect or kill other processes
● Linux kernel allows for nested PID
namespaces
● Processes inside a PID namespace are not
aware of what’s going on outside
● However, processes in the outer PID
namespace consider them as regular
members of the outer process tree

Network Namespaces
● allows a process/group of processes to
see a different set of network interfaces
● each container gets assigned a virtual
network interface
● each virtual network interface is
connected to the Docker daemon’s
docker0 interface
● docker0 routes traffic between containers
and the host (depending on settings)
21

Control groups (cgroups)
● mechnism for limiting certain resources a process/group of processes can
call for
● e.g. CPU, Memory, device access, network (QoS), ..
● a cgroup as a whole can be “frozen” and later “unfrozen”
● freeze mechanism allows to easily stop associated idling processes and to
wake them up if necessary
● might prevent a container from “running amok” (e.g. binds all resources or22

What we’ve learned:
● union file systems enable re-use of single image layers
● a container makes use of CoW in order to work on read-only images
● multiple namespaces provided by host kernel allow for isolated execution
of container processes
● cgroups as a means for limiting access and resource consumption
23

As for security, questions remain:
● The container default user is root. What happens if anyone succeeds
breaking out of a container?
● Is container breakout even possible?
● What about container threats in general?
● What about client-side authentication/authorization in Docker?
● Is there any option to verify the publisher of Docker images in order to
avoid tampering and replay attacks?24

Container security:
How secure is
Docker?
● uid 0 - one account to rule them
all
● Demo: Container breakout
● User namespaces, capabilities
and MAC
● Common container threats
● Docker CLI AuthN/AuthZ
● Docker Content Trust
25

uid 0 - one account to rule them all
● to get that clear: considering the mechanisms introduced so far, there’s
actually no difference between host root and container root!!
● this can even be expanded: Any user allowed to access the Docker
daemon is effectively root on the host!!
● This is also true for otherwise unprivileged users belonging to the docker
group
● Sounds incredible? Watch and be astonished ;)
26

User namespaces to the rescue
● problem: container root is root on host is case of breakout
● solution: “root remapping” (introduced with Docker 1.10)
● maps uid 0 inside container to arbitrary uid outside the container
● caution: user namespaces are disabled by default!
● confinement: there’s only one single user namespace per Docker daemon,
not per container
27

Taming root with Capabilities?
● another problem: setuid-root binaries (e.g. /bin/ping)
● these binaries are also executed with the rights of their owner (guess which
user owns ping)
● heavily increases the risk of privilege escalation in case of flaws
● capabilities idea: grant fine-granular access only to what’s absolutely needed
(network sockets in case of ping)
● allows for unprivileged containers (missing in Docker)28

One last try: Mandatory Access Control
(MAC)
● Linux standard: Discretionary Access
Control (DAC)
● grants access only by the actor’s identity
(access rights per user)
● every resource (file, directory, ..) has a
owning user/group
● Linux manages acess rights for owner,
group and world (rwx)
29
● another approach: Mandatory Access
Control (MAC)
● access is granted by a policy or rather
fine-granular rules
● Linux implementations: SELinux,
AppArmor (rules per file/directory)
● there’re ready-to-use templates offered by
Docker for both
● writing own policy is tricky and error-prone

Container threats: Escaping
● What’s happening?
○ compromising of the host
○ worst case: attacker can do anything on the host system
● Why does it happen?
○ lack of user namespaces
○ insecure defaults/weak configuration (user namespaces disabled)
■ user namespaces disabled30

Container threats: Cross-container attacks
● What’s happening?
○ compromising of sensitive containers (e.g.database container)
○ ARP spoofing and steal of credentials
○ DoS attack (e.g. XML bombs)
● Why does it happen?
○ weak network defaults (default bridge configuration in Docker)
○ poor/missing resource limitation defaults31

Container threats: Inner-container attacks
What’s happening?
attacker gains unauthorized access to a single container
root cause of previous container threats
Why does it happen?
typically due to non-container related flaws (e.g. webapp vulnerabilities)
out of date software
exposing a container to insecure/untrusted networks32

Docker CLI AuthN/AuthZ
[This page is intentionally left blank]
33
Ok,
there’s still TLS ..

Docker Registry: Challenges
● Docker Registry = kind of git repositoy for Docker Images (public/private)
● Handy for collaboration (e.g. in organization context)
➔challenge 1: Make sure that docker pull actually gives us exactly the
content that we want (identity/integrity)
➔challenge 2: Make sure that we always get the latest version of a
requested software (freshness)34

Docker Registry: Attacks (1)
Scenario 1: Attacker hacks into registry server and tampers a single layer of
an up-to-date image (image forgery)
35

Docker Registry: Attacks (2)
Scenario 2: Attacker hacks into registry server and provides content which is
actually out of date (replay attack)
36

Docker Content Trust: Notary
● Implementation of The Update Framework (TUF), which has its origins in
the TOR project (Apache license)
● focus: publisher identity & freshness guarantees
● relies on several keys which are stored at physically different places
● oncept of online and offline keys (compared to simple GPG)
● offline (root) key remains on a USB stick, smart card, ..
37

Docker Content Trust: Notary (2)
38

Docker Content Trust: Registry V2
● Content Addressable System (key-value store)
● pull by hash/pull by digest: key = hash(object)
● self-verifying system (integrity)
● docker pull is a secure operation, as long as we get the correct hash
➔quiz game: How can we ensure to always get the correct hash?
39

Conclusion
● Container systems become more and more security-aware
● However, container security is still work in progress
● Namespaces and Capabilities are relatively new kernel features (buggy?)
● root seems to be a never-ending problem
● Docker provides useful defaults, but lacks support for fine-granular
AuthN/AuthZ
40

Research Questions (1)
Q1: “Will container technology and it’s security become part of a developer’s
everyday life?”
What I think: “Clearly yes! Facing more and more attack vectors every day,
shipping software by means of containers requires at least a basic
understanding of the underlying container system’s security properties. The
DevOps movement relocates responsibilities like deployment, reliability and
security to developers!”
41

Q2: “What about Docker’s future approach in terms of security?”
What I think: “Docker seems to have understood the importance of security for
their tool stack in order to stay successful. In my opinion, the biggest challenge
they’ve to face is integrating security features without damaging the great
usability they offer, since this is what sets them apart from alternative solutions.”
42

Q3: “What about Unikernels? How might this technology help improving Docker
security and Docker in general?”
What I think: “Hard to answer. Regarding one of their blog posts, Docker uses
Unikernels for spawning minimal hypervisors and combines them with the
Docker Engine, creating lightweight apps that contain everything it needs to run
Docker under Non-Linux environments. I’m very excited to hear about their future
plans with Unikernels.”
43

Q4: “Will container systems ever be really secure some day?”
What I think: “In my opinion, there will never be 100% security. The point is: We
saw that containers completely rely on kernel features, they couldn’t even exist
without a kernel. As a consequence, containers will probably be as secure or
rather as unsecure as the operating system they run on.”
44

Thanks for your attention!
Any questions?
Contact:
pk070@hdm-stuttgart.de
Twitter: @Apophis1990
45

Sources (1)
46
Internet:
Docker Inc. (2016): Docker Docs [https://docs.docker.com/]
Docker Inc. (2016): What is Docker? [https://www.docker.com/what-docker]
Ridwan, Mahmud (2016): Separation Anxiety: A Tutorial for Isolating Your System with Linux Namespaces
[https://www.toptal.com/linux/separation-anxiety-isolating-your-system-with-linux-namespaces]
Wikipedia (2016): Descretionary Access Control [https://de.wikipedia.org/wiki/Discretionary_Access_Control]
Wikipedia (2016): Virtuelle Maschine [https://de.wikipedia.org/wiki/Virtuelle_Maschine]
Literature:

Sources (2)
● Graphics:
○ https://assets.toptal.io/uploads/blog/image/674/toptal-blog-image-1416487554032.png
○ https://assets.toptal.io/uploads/blog/image/675/toptal-blog-image-1416487605202.png
○ https://c2.staticflickr.com/8/7336/14098888813_1047e39f08.jpg
○ https://blog.docker.com/wp-content/uploads/2015/08/dct1.png
○ https://docs.docker.com/engine/article-img/architecture.svg
○ https://docs.docker.com/engine/userguide/storagedriver/images/aufs_delete.jpg
○ https://docs.docker.com/engine/userguide/storagedriver/images/aufs_layers.jpg
○ https://www.docker.com/sites/default/files/what-is-docker-diagram.png
○ https://www.docker.com/sites/default/files/what-is-vm-diagram.png
○ https://www.docker.com/sites/default/files/products/what_is_layered_filesystems_sm.png
47

Exploring Docker Security

More Related Content

Exploring Docker Security

Editor's Notes