The information in this document focuses primarily on cloud based deployments. For on-premise deployments, additional security work that is specific to your installation method would also be required. Note that your specific installation’s security needs might be more or less stringent than what we can offer you here.
Brad Geesamen gave a wonderful talk titled Hacking and Hardening Kubernetes by Example at Kubecon NA 2017. You can watch the talk or read the slides. Highly recommended that you do so to understand the security issues you are up against when using Kubernetes to run JupyterHub.
If you find a security vulnerability in JupyterHub, either a failure of the
code to properly implement the model described here, or a failure of the
model itself, please report it to email@example.com.
If you prefer to encrypt your security reports, you can use
this PGP public key.
This section describes how to enable HTTPS on your JupyterHub. The easiest way to do so is by using Let’s Encrypt, though we’ll also cover how to set up your own HTTPS credentials. For more information
on HTTPS security see the certificates section of this blog post.
Buy a domain name from a registrar. Pick whichever one you want.
Create an A record from the domain you want to use, pointing to the EXTERNAL-IP of the proxy-public service. The exact way to do this will depend on the DNS provider that you’re using.
Wait for the change to propagate. Propagation can take several minutes to several hours. Wait until you can type in the name of the domain you bought and it shows you the JupyterHub landing page.
It is important that you wait - prematurely going to the next step might cause problems!
JupyterHub uses Let’s Encrypt to automatically create
HTTPS certificates for your deployment. This will cause your HTTPS certificate
to automatically renew every few months. To enable this, make the following
changes to your config.yaml file:
Specify the two bits of information that we need to automatically provision
HTTPS certificates - your domain name & a contact email address.
Apply the config changes by running helm upgrade ...
helm upgrade ...
Wait for about a minute, now your hub should be HTTPS enabled!
If the proxy service is of type LoadBalancer, which it is by default, then a specific static IP address can be requested (if available) instead of a dynamically acquired one.
Although not essential for HTTPS, using a static IP address is a recommended practice for domain names referencing fixed IPs.
This ensures the same IP address for multiple deployments.
The IP can be provided like:
More info about this can be found on the Configuration Reference page.
If you have your own HTTPS certificates & want to use those instead of the automatically provisioned Let’s Encrypt ones, that’s also possible. Note that this is considered an advanced option, so we recommend not doing it unless you have good reasons.
There are two ways to specify your manual certificate, directly in the config.yaml or by creating a Kubernetes secret.
Add your domain name & HTTPS certificate info to your config.yaml
-----BEGIN RSA PRIVATE KEY-----
-----END RSA PRIVATE KEY-----
Apply the config changes by running helm upgrade ….
Create a secret resource with type kubernetes.io/tls containing your certificate.
kubectl create secret tls example-tls --key="tls.key" --cert="tls.crt"
Add your domain and the name of your secret to your config.yaml.
In some environments with a trusted network, you may want to terminate SSL at a
load balancer. If https is enabled, and proxy.https.type is set to offload,
the HTTP and HTTPS front ends target the HTTP port from JupyterHub.
The HTTPS listener on the load balancer will need to be configured based on the
provider. If you’re using AWS and a certificate provided by their certificate
manager, your config.yml might look something like:
# Certificate ARN
# The protocol to use on the backend, we use TCP since we're using websockets
# Which ports should use SSL
Annotation options will vary by provider.
There are many ways to confirm that a domain is running trusted HTTPS
certificates. One options is to use the Qualys SSL Labs
security report generator. Use the following URL structure to test your domain:
Helm 3 supports the security, identity, and authorization features of modern Kubernetes. Helm’s permissions are evaluated using your kubeconfig file. Cluster administrators can restrict user permissions at whatever granularity they see fit.
Read more about organizing cluster access using kubeconfig files in the
Most cloud providers have a static IP you can hit from any of the compute nodes, including the user pod, to get metadata about the cloud. This metadata can contain very sensitive info, and this metadata, in the wrong hands, can allow attackers to take full control of your cluster and cloud resources. It is critical to secure the metadata service. We block access to this IP by default (as of v0.6), so you are protected from this!
The slides beginning at Slide 38 provides more information on the dangers presented by this attack.
If you need to enable access to the metadata server for some reason, you can do the following in config.yaml:
The Kubernetes Dashboard gets created by default in many installations. Although the Dashboard contains useful information, the Dashboard also poses a security risk. We recommend deleting it and not using it for the time being until the Dashboard becomes properly securable.
You can mitigate this by deleting the Kubernetes Dashboard deployment from your cluster. This can be most likely performed with:
kubectl --namespace=kube-system delete deployment kubernetes-dashboard
In older clusters, you might have to do:
kubectl --namespace=kube-system delete rc kubernetes-dashboard
Kubernetes supports, and often requires, using Role Based Access Control (RBAC)
to secure which pods / users can perform what kinds of actions on the cluster. RBAC rules can be set to provide users with minimal necessary access based on their administrative needs.
It is critical to understand that if RBAC is disabled, all pods are given root equivalent permission on the Kubernetes cluster and all the nodes in it. This opens up very bad vulnerabilites for your security.
As of the Helm chart v0.5 used with JupyterHub and BinderHub, the helm chart can natively work with RBAC enabled clusters. To provide sensible security defaults, we ship appropriate minimal RBAC rules for the various components we use. We highly recommend using these minimal or more restrictive RBAC rules.
If you want to disable the RBAC rules, for whatever reason, you can do so with the following snippet in your config.yaml:
We strongly discourage disabling the RBAC rules and remind you that this
action will open up security vulnerabilities. However, some cloud providers may
not support RBAC in which case you can disable it with this config.
Allowing direct user access to the Kubernetes API can be dangerous. It allows
users to grant themselves more privileges, access other users’ content without
permission, run (unprofitable) bitcoin mining operations & various other
not-legitimate activities. By default, we do not allow access to the service
account credentials needed
to access the Kubernetes API from user servers for this reason.
If you want to (carefully!) give access to the Kubernetes API to your users, you
can do so with the following in your config.yaml:
You can either manually create a service account for use by your users and
specify the name of that here (recommended) or use default to give them access
to the default service account for the namespace. You should ideally also
(manually) set up RBAC
rules for this service account to specify what permissions users will have.
This is a sensitive security issue (similar to writing sudo rules in a
traditional computing environment), so be very careful.
There’s ongoing work on making this easier!
Kubernetes has optional support for network
which lets you restrict how pods can communicate with each other and the outside
world. This can provide additional security within JupyterHub, and can also be
used to limit network access for users of JupyterHub.
By default, the JupyterHub helm chart disables network policies.
Important: If you decide to enable network policies, you should be aware
that a Kubernetes cluster may have partial, full, or no support for network
policies. Kubernetes will silently ignore policies that aren’t supported.
Please use caution if enabling network policies and verify the policies
behave as expected, especially if you rely on them to restrict what users can
You can enable network policies in your config.yaml:
The default singleuser policy allows all outbound network traffic, meaning
JupyterHub users are able to connect to all resources inside and outside your
network. To restrict outbound traffic to DNS, HTTP and HTTPS:
- port: 53
- port: 80
- port: 443
See the Kubernetes
for further information on defining policies.
By default any IP address can access your JupyterHub deployment through the load balancer service.
In case you want to restrict which IP addresses are allowed to access the load balancer, you can
specify a list of IP CIDR addresses in your config.yaml as follows:
This would restrict the access to only two IP addresses: 22.214.171.124 and 126.96.36.199.