Best practices in Kubernetes security are rapidly evolving. Many security problems in early versions of Kubernetes are resolved by default in recent versions. However, like any complex system, there are still risks you should understand before you trust it with your production data. We’ve tried to summarize the most important things you should have in mind when you host sensitive workloads on Kubernetes.
The topics discussed here help you understand potential risks in your cluster. The risk in your environment depends on your threat model and the types of applications you run in your cluster. You’ll have to consider how best to invest in security controls and hardening based on the sensitivity of your data, the amount of time and staff you’re able to dedicate to security, and your company’s particular compliance requirements.
Kubernetes provides several mechanisms to enforce security within the cluster. These range from API security controls, down to container isolation, resource limiting, and network policy control.
Kubernetes core components cooperate to schedule and run your workloads in a cluster. Kubernetes provides a range of access control mechanisms, however their default values tend to be overly permissive. You should carefully determine what access your system components and users need, and configure the most restrictive controls possible.
Remember also to secure the infrastructure that your clusters run on - for example, SSH access, or cloud provider access such as AWS IAM.
Kubernetes clusters require PKI certificates for secure communication between cluster components. Default CAs and certificates are provided by kubeadm, but you should consider your requirements before accepting only the defaults.
Kubernetes requires TLS for the communication between the control plane components of your cluster. For details about the required PKI certificates, see the certificates documentation. You can reuse the control plane CA certificate bundle for TLS in your application/workloads. See the documentation about managing TLS certificates. kubeadm automatically generates the certificates required by the cluster; this topic explains when and why you might want to generate your own certificates. It also discusses options for managing certificates for your applications.
The certificates generated with a kubeadm install rely on a single cluster CA for all certificates. You might want to manage your certificates differently in the following cases:
- For finer-grained control over authentication, you might set up different certificate authorities for server certificates and client certificates.
- Your company’s policies might require TLS certificates that are issued by your own PKI.
- If you integrate with an OpenID Connect (OIDC) provider, you can use the OIDC CA.
- Publicly facing workloads may require a Commercial or non-profit certificate bundle.
If you issue your own certificates
The certificate for the API server control plane component requires a
kubernetes. We recommend you use a corporate CA on a load balancer in front of
the API server instead of replacing the CA. Note that the
for the controller manager must also include a copy of the CA for the
By default, kubeadm generates certificate authorities that expire after 10 years. The server and client certificates expire after one year. It is strongly recommended to not let certificates expire as your cluster will become inoperable.
To help make sure your certificates do not expire, you can use the Prometheus BlackBox exporter, which allows probing and alerting on certificate expiry dates.
Kubernetes uses client certificates, bearer tokens, an authenticating proxy, or HTTP basic auth to authenticate API requests through authentication plugins. As HTTP requests are made to the API server, plugins attempt to associate the following attributes with the request: Username, UID, Groups, and Extra fields.
A critical component of cluster security is making sure that human users, Kubernetes services accounts, cluster components, and application components have the right permissions to access only the resources they need to get their respective jobs done. Authentication and authorization are critical parts of access control.
Integrate an identity provider
Certificates take care of authentication for clients, servers, clusters and applications. To authenticate human users, we recommend integrating an existing corporate identity system. Kubernetes lets you provide authentication with any compliant OpenID Connect provider (for example, GitHub or Google). Kubernetes authentication and authorization can also be extended with webhook-based plugins to create a custom identity integration.
If your organization integrates multiple identity providers,
Dex can be integrated with Gangway to act as
the OIDC endpoint. Dex acts as a broker for identity, providing a standard OIDC
frontend for a variety of backends such as LDAP servers, SAML providers, or
established identity providers like GitHub, Google, and Active Directory.
Multi-factor authentication is not required, but provides additional protection for end user flows.
Human users and service accounts need to be carefully authorized to access only the resources they need to get their jobs done, and no more. The principle of least privilege is central to good authorization policies.
Kubernetes expects attributes that are common to REST API requests. This means that Kubernetes authorization works with existing organization-wide or cloud-provider-wide access control systems, which may handle other APIs besides the Kubernetes API. Every authenticated request to the Kubernetes API server made by a human user or a service account needs to be authorized.
Role-Based Access Control (RBAC) allows the control of actions performed on resources in the cluster, and defines who is allowed to perform them. Every resource in Kubernetes is represented as an API object (Pods, Namespaces, Secrets, ConfigMaps, etc.) These resources can be created, read, updated, and deleted (verbs). A rule is composed of a verb and a resource, as an operation to be performed on an API group. These rules are bundled together in Roles. Roles are scoped to a namespace. Cluster-wide roles are defined in ClusterRole objects. Roles can then be bound to users, groups and service accounts by creating a role binding thereby granting them the ability to perform actions described in the roles.
At a minimum, we recommend that you enable RBAC. RBAC is enabled by default in most recent installers and provides a framework for implementing the principle of least privilege for humans and applications that access the Kubernetes API.
To get the most benefit from RBAC, an appropriate configuration is required:
- Run each component with the most restrictive permissions that still allow for expected functionality. Most applications in a cluster will need little or no access to the Kubernetes API. System components such as an ingress controller or monitoring system may need more access, but can often be limited to read-only access or access within a particular namespace.
- Make sure that trusted components don’t act as pivots that allow less privileged users to escalate privileges. The Kubernetes Dashboard and Helm tiller daemon are examples that deserve special attention. Isolate these components with application-level authentication/ authorization and network access controls to prevent unauthorized access.
When creating RBAC policies, prefer Roles and RoleBindings over ClusterRoles and ClusterRoleBindings whenever possible as they are scoped to namespaces by default. While Kubernetes comes with default RBAC policies in place, we recommend setting up your baseline policies with the least required privileges that you need.
The Kubernetes API audit logs are a useful tool for discovering which APIs a particular application is using, and for testing locked-down RBAC policies. The audit2rbac tool can act as a reference as it can generate RBAC roles and bindings to cover the API requests made by a user. See also Auditing in the Kubernetes documentation.
An admission controller is a piece of code that intercepts requests to the Kubernetes API. This happens before the persistence of the object, but after the request is authenticated and authorized. These are plugins that govern and enforce the acceptance of requests. There are two individual Admission Controllers: MutatingAdmissionWebhook and ValidatingAdmissionWebhook, which execute mutating and validating actions, respectively. Mutating controllers may modify the objects they admit; validating controllers do not. The admission control process has two phases: the mutating phase is executed first, followed by the validating phase. An excellent example of a mutating admission controller is Istio’s automatic sidecar injection mechanism. If any of the controllers in either phase reject the request, the entire request is rejected immediately and an error is returned to the end-user
Network and Application Access Control
Access to the cluster network should be carefully controlled and permissions granted only to the components or resources that need access. The Kubernetes NetworkPolicy API allows users to express ingress and egress policies (starting with Kubernetes 1.8.0) to Kubernetes pods based on labels and ports.
Many existing applications assume that network-level access implies a level of authorization. Even if applications include strong application-layer authentication and authorization, network-level access control provides an additional layer of defense. For example, it provides crucial protection against pre-auth vulnerabilities such as the Heartbleed (CVE-2014-0160) vulnerability in OpenSSL.
By default, Kubernetes clusters do not restrict traffic. Pods can communicate with any other pods. External clients can also communicate with pods, assuming they are routable from the client’s network.
The NetworkPolicy resource in Kubernetes allows you to control how pods are allowed to communicate with each other and other network endpoints. The NetworkPolicy resource is namespace scoped. Rules defined in the policy allow traffic and are combined additively.
Kubernetes provides core data types for specifying network access controls between pods. Network policy in Kubernetes can limit inbound traffic to a pod based on the source pod’s namespace and labels, plus the IP address for traffic that originates outside the cluster. Network policy can also limit outbound traffic using the same set of selectors. A good starting point is to restrict ingress to only the application namespace by default. For details, see the Kubernetes Network Policy documentation.
The enforcement of network policy relies on the cluster’s CNI provider. Without them, Kubernetes “fails open” — the API happily accepts any network policy, but the policies are not enforced. We recommend Calico as your CNI provider, because it enforces controls. Examples of Network policies can be found here.
Restricting access to control plane services
Network controls in the infrastructure underlying the cluster must also be considered. In a cloud provider environment, make sure that pods cannot communicate with the instance metadata service. We also recommend the use of the Node Authorizer to limit kubelet access to the API. When enabled, this special-purpose authorization module restricts kubelet access to resources that are referenced by Pods running on that specific node. The Node Authorizer is enabled by default in recent releases of kubeadm. For example, instead of being able to access all Secrets in the cluster, a kubelet can access only Secrets that are referenced by Pods scheduled to that kubelet.
Enforcing network controls in the infrastructure underlying your cluster is also
critical. In a cloud provider environment, make sure that your pods cannot talk
to the instance metadata service (for example,
http://169.254.169.254/latest/meta-data on AWS EC2). Depending on your
requirements, you may also need to restrict access to the kubelet localhost
read-only port (10255 by default). This port exposes metadata about the pods
running on the node, which you may not want access to your applications.
Application-layer access control
One solution to the problem of network-level access controls is strong application-layer authentication such as mutual TLS. Cryptographic application identity is powerful because it allows identity to be efficiently expressed across network boundaries. Securely provisioning certificates for applications is still a hard problem in Kubernetes.
Network access controls have some limitations in dynamic environments like Kubernetes, which results in the following difficulties:
- Federating Kubernetes network policy across multiple clusters.
- Integrating Kubernetes network-level controls and granular network-level controls expressed outside of the pod networking layer (for example, in AWS EC2 Security Groups).
- When services running in a Kubernetes cluster need to communicate with services outside the cluster, NetworkPolicy is often unable to filter traffic as expected due to source and destination IP address translation.
If you encounter any of these issues, we recommend that you define a more coarse-grained network policy and rely on the application layer for fine-grained access control.
Security contexts limit what a Pod or Container can do and what privileges the object has when running in the cluster. Example controls are the UID of the process running inside the container, the filesystem access group, the process capabilities, SELinux labels, etc.
These can be applied to individual Pods and containers, and they define a set of conditions that it must run with.
Pod security policies are cluster-wide resources that provide automation of the
above described security contexts. PSPs can be used to automatically set
security context parameters or to prevent out-of-policy pods from running in the
cluster. For example, if you don’t want any containers in your cluster to run as
root, you can enforce this using a PSP with a
runAsUser rule of
MustRunAsNonRoot. They define a set of conditions that a pod must run with to
be accepted into the system. Pod Security Policies are comprised of settings and
strategies that control the security features a pod has access to, and hence
this must be used to control pod access permissions. Strong pod security
policies make sure that pod access is appropriately controlled.
Pod Security Policies provide a policy-driven mechanism for requiring applications in your cluster to use container sandboxing in an approved way. For example, you can require that all pods in a particular namespace run as non-root, that they don’t mount host file systems and do not use host networking.
To use pod security policies, the PodSecurityPolicy admission controller must be enabled in the API server configuration. Policies must be present before enabling the controller, or no pods would be allowed to run. For more information, see to the Kubernetes documentation.
Credentials security (Secrets)
Secrets are sensitive pieces of data such as passwords, tokens or keys. Applications use secrets to access internal resources like the Kubernetes API or external resources such as git repositories, databases, etc. The following section details concerns related to secrets in the context of Kubernetes.
Kubernetes has a core primitive for managing application secrets, appropriately called a Secret. Applications typically need secrets for two reasons:
- They need access to a credential that proves their identity to another system (for example, a database password or third-party API token).
- They need a cryptographic secret for some essential operation (for example, an HMAC signing key for issuing signed HTTP cookies).
For the first use case of application identity, follow the efforts of SPIFFE and the Container Identity working group for a long term solution to dynamically provisioning unique application identities. In the near term, there is no well-established best practice in this area. Still, some users have success integrating with existing certificate provisioning workflows as part of a CI/CD pipeline. Simple Kubernetes-native solutions like cert-manager may also work for your use case.
For the other use case, systems such as Vault perform cryptographic operations in a centralized service. If you choose this option, make sure you understand the entire chain of attestations involved in authenticating to the system. Often these systems depend on Kubernetes secret resources as one step in the chain. In Vault, the Vault Kubernetes auth backend authenticates pods by consuming a Kubernetes Service Account token. Still, the token is stored as a secret object before it’s injected into the pod. This pattern requires that you trust Vault not to replay your token and impersonate the pod to the Kubernetes API.
Caveats for Kubernetes Secrets
You should be aware of the following limitations:
- Many standard components – for example, ingress controllers – require permission to read all secrets in your cluster.
- Secrets are not encrypted at rest by default. You can, however, configure, etcd, to encrypt secret data at rest. For details, see Encrypting Secret Data at Rest.
Auditing to support security
Audit logging must be explicitly enabled. It provides valuable insight into access rules, compliance, and potential access issues. Kubernetes auditing provides a security-relevant chronological set of records documenting the sequence of activities that have affected the system by individual users, administrators, or other components of the system.
The Kubernetes API server audit log documents the sequence of cluster activities performed by users, administrators, and system services. Each API request has multiple stages that can be tracked and logged using an Audit Policy.
Enabling Audit Logging
The Kubernetes API server does not perform audit logging by default. We recommend that you enable audit logging to a file by setting the following flags in the API server configuration:
--audit-log-pathspecifies the log file path that log backend uses to write audit events.
--audit-log-maxagedefines the maximum number of days to retain old audit log files
--audit-log-maxbackupdefines the maximum number of audit log files to retain
--audit-log-maxsizedefines the maximum size in megabytes of the audit log file before it gets rotated
--audit-policy-filespecifies the Audit policy file to be used
Audit events can also be sent to a webhook backend, but we recommend logging to a file that can be aggregated. Audit data should be treated as a high priority, and care should be taken that the file specified for –audit-log-path is aggregated for multiple control plane nodes and handled by systems with high reliability. In the case of an outage or other issue, administrators need to be able to rely on the data produced by audit systems.
When logging to files on the control plane hosts, you should set
appropriately based on the available disk space before aggregation.
Node and container runtime hardening
It is of critical importance to consider the security of the container-host boundary. This is important even in single-tenant environments since a remote code execution vulnerability like Shellshock (CVE-2014-6271) or the Ruby YAML parsing vulnerability (CVE-2013-0156) can turn your otherwise trusted workload into a malicious agent. Without proper hardening, that single remote code execution vulnerability can escalate into a whole-node or whole-cluster takeover.
Current container runtimes don’t provide the most reliable possible sandboxing, but there are some steps you can take to help mitigate the risk of container escape vulnerabilities:
- Segment your Kubernetes clusters by integrity level — a simple but very effective way to limit your exposure to container escape vulnerabilities. For example, your dev/test environments might be hosted in a different cluster than your production environment.
- Invest in streamlined host/kernel patching. Make sure that you have a way to test new system updates (for example, a staging environment) and that your applications can tolerate a rolling upgrade of the cluster without affecting application availability.
- Kubernetes shines at orchestrating these upgrades. Once you build confidence in letting Kubernetes dynamically rebalance application pods, patch management at the node level becomes relatively easy. You can automate a rolling upgrade that gracefully drains each node and either upgrade it in place or (in an IaaS environment) replaces it with a new node. Investments in this area also improve your overall resiliency to node-level outages.
- Run your applications as a non-root user. Root (UID 0) in a Linux container is still the same user as root on the node. A combination of sandboxing mechanisms restrict what code running in the container can do. Still, future Linux kernel vulnerabilities are more likely to be exploitable by a root user than by a non-privileged user.
- Enable and configure extra Linux security modules like SELinux and AppArmor. These tools let you enforce more restrictive sandboxing on particular containers. They are valuable in many situations, but building and maintaining appropriate configurations requires a time investment. They may not be appropriate for every application or environment.
Runtime security is concerned with potential changes to a running container through its lifetime, invalidating an initial security scan. A container image could have been scanned and approved but become a liability as new vulnerabilities, bugs, and threats are found. Runtime security tools help to mitigate this problem by looking at what’s happening inside containers: filesystem, process activity, networking behavior, etc. Examples of runtime security tools: Falco, Aquasec, Twistlock, Sysdig.
Attack Surface Minimization
Minimize container footprint and attack surface by excluding extraneous libraries and utilities that are not needed and could be leveraged during an attack. Consider building images from scratch and include only what is necessary at runtime. Also leverage multi-stage builds where applicable so that build tools are not included in the final image used in production.
Container Image Scanning
Container image scanning is an integral part of building container images, whether from source or third party base images, to discover any known vulnerabilities and mitigate them before cluster deployment. One of the last steps of your CI (Continuous Integration) pipeline involves building the container images that would be pulled and executed in your environment. Therefore, whether you are building Docker images from your code or using unmodified third party images, it’s crucial to identify and find any known vulnerabilities that may be present in those images. This process is known as container image scanning. Container image scanning helps make sure that your images are free from known vulnerabilities before they are deployed. Appropriately managed scanning keeps your clusters safe by not introducing malicious or malformed artifacts.
Services and open source tools that provide image scanning include the following:
We recommend integrating a container scan as part of a continuous delivery pipeline. In addition, we recommend running periodic scans against stored images so you can identify and mitigate new vulnerabilities.
We also recommend deploying an ImagePolicyWebhook admission controller with the Kubernetes API server that only allows images that have passed security scans to run in a cluster.
Deploying the ImagePolicyWebhook in Kubernetes
The ImagePolicyWebhook admission controller plugin queries a backend service to determine whether a workload can be run on a cluster. It does not make any changes to the submitted workload, but instead accepts or rejects it as-is based on whether the images associated with the workload comply with the policy set for the cluster. The webhook service queries the scanner tool to scan an image and either accept the workload or reject it based on the scan results.
The webhook service must use TLS.
See the documentation on ImagePolicyWebhook for a detailed explanation.
Image signing helps make sure that your container images have not been tampered with. In other words, image signing is used to prove the provenance of your images. Signed images do not guarantee compliance, however.
Image signing establishes image trust, ensuring that the image you run in your cluster is the image you intended to run.
Private image registries or private accounts on public registries let you establish some degree of trust. Still, if your registry is compromised, bad actors could replace your images with malicious versions.
Image signing adds a layer of protection by cryptographically signing an image. As long as your private keys are not compromised, you can be guaranteed that the image you run is trusted.
Note that you must also deploy a mechanism to verify image trust. An example is portieris, which allows you to configure image security policies and stop the workload from being deployed if it is not signed.
Patch management and CI/CD
A successful cluster access pattern is to have most users interact with the production cluster only through a deployment pipeline. This pipeline consists of one or more automated systems that handle building code into a container image, running unit and integration tests and other validation steps such as pausing for any manual approval. Depending on your needs, developers could still have direct read-only access to the Kubernetes API or have a way to “break the glass” and exec into pods during an incident.
A robust application deployment pipeline is also the key to remediating vulnerabilities in container images. You can use tools like Clair to identify known vulnerabilities in the libraries and packages you use. Still, to release patches on time, you need a trusted, automated way of rebuilding and testing patched versions of the container.
Healthy Kubernetes clusters are dynamic environments. New versions of applications are deployed, nodes disappear for kernel upgrades, deployments scale up and down, and (hopefully) the users of your application never notice. Making all this work in practice requires some diligence, but it’s critical to reaping all the benefits of Kubernetes.
One tool that can help put bounds on the amount of chaos introduced into your cluster is the Pod Disruption Budget. It’s useful when you have multiple automated systems, and you want to make sure they don’t interact in unwanted ways. For example, an application-level bug might leave some pods of your application temporarily unavailable. A pod disruption budget could make sure that an automated rolling node upgrade doesn’t terminate the remaining healthy copies of your application.
Overly privileged container builds
One Docker-specific anti-pattern to avoid in your build pipeline is mounting the host-level Docker control socket “/var/run/docker.sock” into a container during a build. Access to this socket is equivalent to root on the host, which means any running build could compromise the node. This is doubly true if your build system runs build before a manual code review (a typical pattern).
What to do now
We keep saying it: how you secure your Kubernetes cluster depends in part on your available resources and your application requirements. Consider each element in the broader security picture and spend some time upfront assessing how important it is to your needs overall. At a very high level, some of our recommendations fit nicely into larger best practices in deployment:
- Automated deployment pipeline and scheduler. Lets you simplify host and application patch management with rolling upgrades that are integrated into the rest of your overall development cycle.
- Integrated access controls at appropriate levels. (authz/authn with API integration)
- Integrated logging and monitoring. You log and monitor for performance and reliability – adding support for security-specific events and pod metadata is non-trivial but vital. Precisely what to monitor depends as always on your specific needs.
Planning for the future
Security is an increasing concern for everyone, and initiatives are well underway to improve the security landscape. Keep an eye out for developments on these fronts:
- More strongly encrypted identity specific to your hardware or cloud provider
- Stronger provenance for cryptographically signed binaries/images
- Automatically updated inventories
- More sophisticated alerts and monitoring
- Securing a Cluster (Kubernetes documentation)
- Hacking and Hardening Kubernetes By Example (slides from Brad Geesaman)
Popular Tooling and Approaches
The Kubernetes docs provide instructions to enable encryption at rest for specified resources. Most importantly, this allows cluster operators to ensure the data contained in Secrets has a layer of protection against a malicious actor gaining access to the storage disk for etcd.
- Adds a layer of protection for data contained in Secrets
- Is compromised if attacker gains access to the encryption keys used by the API Server
Validating Admission Webhook Controllers
Kubernetes RBAC provides a useful way to manage access to resources. However, it
does not provide ways to restrict access based on external systems or the
attributes of the resource being created. Proper authorization may require a
deeper understanding of the requester, evaluation of business logic, or
understanding of the cluster’s current state to determine whether a request
should be authorized. Kubernetes offers preset admission controllers built into
kube-apiserver, such as
PodSecurityPolicy. These can be enabled by
altering flags on the
kube-apiserver. To provide custom admission control
without modifying the API server, Kubernetes offers an Admission webhook
functionality. In this model, Kubernetes offers selected inbound request to an
external service that can approve or deny the request.
Writing an admission webhook controller in code enables you to do complex logic and access Kubernetes objects structurally, through a language’s type system. For example, see https://github.com/kubernetes/apimachinery. A downside to this approach is maintaining the controller code base over time, which can involve keeping logic up to date against changing Kubernetes API versions.
- Flexible option for resource validation
- Webhook is on the critical path for resource management in the cluster
- Development & maintenance overhead for the webhook
Open Policy Agent (OPA)
Similar to Validating Admission Webhook Controllers, OPA performs validation on requests sent to it. OPA is a general purpose policy agent. It’s primary goal is to unify policies around a centralized model and language. It uses a DSL called Rego that analyzes input, usually JSON, and provides an output. For Kubernetes admission control, the request is typically the JSON from an AdmissionReview object and the response is the same object with the AdmissionResponse filled in.
While this unified model is great, it can be harder to do complex logic in rego over a general purpose language.
- Powerful validation framework
- Requires learning a new policy definition language: Rego
Kubernetes provides a NetworkPolicy API. Verify that your CNI-Plugin enforces policies. By default, Kubernetes allows traffic to and from any pod or external source that can reach it. With this in mind, enforcing Network Policy is critical to protect applications from unwanted access. How policy is enforced depends on the CNI-plugin. For example, in Calico it’s enforced via IPtables and in Cilium it’s BPF. If network policy if a large part of your Kubernetes design, ensure the solution you’re using offers a scalable approach to enforcement.
Along with the Kubernetes network policy API, some CNI-Plugins offer their own CRDs that extend the base functionality. Calico offers a GlobalNetworkPolicy and NetworkPolicy CRD. The primary trade-off to using a CNI-specific CRD is portability. Should you choose to change plugins in the future, you may need to convert policies between types. However, some of the plugin-specific policy features are very compelling for various cluster architectures.
Historically, organizations have attempted to keep their existing network models outside of Kubernetes and make Kubernetes fit within it. For example, this can include making layers of Kubernetes nodes such as running nodes in Kubernetes dedicated to certain layers such as Web, App, and Data. This moves away from Kubernetes declarative approach and add a lot of complexity to the system. We highly recommend considering intra-cluster and workload traffic policies into the Kubernetes API.
- Essential network controls can be enforced
- Differences in implementation between CNI plugins can be a challenge