k-rail is a workload policy enforcement tool for Kubernetes. It can help you secure a multi tenant cluster with minimal disruption and maximum velocity.
- Why k-rail?
- Installation
- Removal
- Viewing policy violations
- Supported policies
- No ShareProcessNamespace
- No Exec
- No Bind Mounts
- No Docker Sock Mount
- EmptyDir size limit
- Mutate Default Seccomp Profile
- Immutable Image Reference
- No Host Network
- No Host PID
- No New Capabilities
- No Privileged Container
- No Helm Tiller
- Trusted Image Repository
- Safe to Evict (DEPRECATED)
- Mutate Safe to Evict
- Mutate Image Pull Policy
- Require Ingress Exemption
- Unique Ingress Host
- Configuration
- Adding new policies
- Debugging
- License
By default, the Kubernetes APIs allow for a variety of easy privilege escalation routes. When operating a multi-tenant cluster, many features can be dangerous or introduce instability and must be used judiciously. k-rail attempts to make workload policy enforcement easy in Kubernetes, even if you already have a large number of diverse workloads. Several features enable you to roll out policy enforcement safely without breaking existing workloads:
- Passive report-only mode of running policies
- Structured violation data logged, ready for analysis and dashboards
{ "enforced": true, "kind": "PodExec", "namespace": "ecommerce", "policy": "pod_no_exec", "resource": "payment-processor", "user": "bob@amyshardware.com", "time": "2019-11-03T06:28:07Z" }
- Flexible and powerful policy exemptions by resource name, namespace, groups, and users
--- - resource_name: "*" namespace: "cluster-conformance-testing" username: "cluster-ci@paas-ci.iam.gserviceaccount.com" group: "*" exempt_policies: - "pod_no_privileged_containers" - "pod_no_bind_mounts" - "pod_no_host_network" - "pod_default_seccomp_policy" - "pod_no_host_pid" - "pod_no_exec"
- Realtime interactive feedback for engineers and systems that apply resources
$ kubectl apply --namespace default -f deploy/non-compliant-ingress.yaml Error from server (InternalError): error when creating "deploy/non-compliant-ingress.yaml": Internal error occurred: admission webhook "k-rail.cruise-automation.github.com" denied the request: Ingress bad-ingress had violation: Require Ingress Exemption: Using the 'public' Ingress class requires an exemption Ingress bad-ingress had violation: Requires Unique Ingress Host: Ingress Host should not point to multiple namespaces
By leveraging the first three features you can quickly and easily roll out enforcement to deployments without breaking them and monitor violations with confidence. The interactive feedback informs and educates engineers during future policy violations.
Cruise was able to utilize this software to apply enforcement to more than a dozen clusters with thousands of existing diverse workloads in all environments in about a week without breaking existing deployments. Now you can too.
If you have a new cluster without existing workloads, just run k-rail in enforcement mode for the policies you desire and add exemptions as needed.
If you have a cluster with existing workloads, run it in monitor mode for a few weeks or until you have collected enough data. The violation events are emmitted in the logs in JSON, so it is suggested that you analyze that data collected to make exemptions as needed. Once the exemptions are applied, you can safely turn on enforcement mode without breaking existing workloads.
You can install or update k-rail using the helm chart in deploy/helm
.
For the Helm deployment, all configuration for policies and exemptions are contained in deploy/helm/values.yaml
.
For Helm 2 and below, it is recommended to use helm template
render the YAML for applying rather than using Helm Tiller:
helm template --namespace k-rail deploy/helm | kubectl apply -f -
By default all policies are enforced (report_only: false
).
Test the default configuration by applying the provided non-compliant deployment:
kubectl apply --namespace default -f deploy/non-compliant-deployment.yaml
Removing k-rail is similar to the installation, but you use delete
instead of apply
:
helm template --namespace k-rail deploy/helm | kubectl delete -f -
There are a few ways of viewing violations. Violations from realtime feedback and the Events API is most useful for users, but violations from logs is most useful for presentation and analysis.
You may see violations when applying your resources:
$ kubectl apply -f deploy/non-compliant-deployment.yaml
Error from server (k-rail): error when creating "deploy/non-compliant-deployment.yaml": admission webhook "k-rail.cruise-automation.github.com" denied the request:
Deployment bad-deployment had violation: Host Bind Mounts: host bind mounts are forbidden
Deployment bad-deployment had violation: Docker Sock Mount: mounting the Docker socket is forbidden
Deployment bad-deployment had violation: Immutable Image Reference: image tag must include its sha256 digest
Deployment bad-deployment had violation: No Host Network: Using the host network is forbidden
Deployment bad-deployment had violation: No Privileged Container: Using privileged containers is forbidden
Deployment bad-deployment had violation: No New Capabilities: Adding additional capabilities is forbidden
Deployment bad-deployment had violation: No Host PID: Using the host PID namespace is forbidden
Deployment bad-deployment had violation: Safe to evict: annotation is required for Pods that use emptyDir or hostPath mounts to enable cluster autoscaling
You can also see violations that have occurred recently with the events API:
$ kubectl get events --namespace default
LAST SEEN TYPE REASON KIND MESSAGE
3m41s Warning FailedCreate ReplicaSet Error creating: admission webhook "k-rail.cruise-automation.github.com" denied the request:
bad-pod-5f7cd9bf45-rbhsb had violation: Docker Sock Mount: mounting the Docker socket is forbidden
Violations are also emitted as structured data in the logs:
$ kubectl logs --namespace k-rail --selector name=k-rail | jq '.'
{
"enforced": true,
"kind": "Deployment",
"namespace": "default",
"policy": "pod_no_host_network",
"resource": "evil-deployment",
"time": "2019-10-23T19:54:24Z",
"user": "dustin.decker@getcruise.com"
}
{
"enforced": true,
"kind": "Deployment",
"namespace": "default",
"policy": "pod_no_privileged_container",
"resource": "evil-deployment",
"time": "2019-10-23T19:54:24Z",
"user": "dustin.decker@getcruise.com"
}
{
"enforced": true,
"kind": "Deployment",
"namespace": "default",
"policy": "pod_no_new_capabilities",
"resource": "evil-deployment",
"time": "2019-10-23T19:54:24Z",
"user": "dustin.decker@getcruise.com"
}
Since the violations are outputted as structured data, you are encouraged to aggregate and display that information. GCP BigQuery + Data Studio, Sumologic, Elasticsearch + Kibana, Splunk, etc are all capable of this.
shareProcessNamespace: true
is a Pod Spec directive that puts all containers in a Pod within
the same PID Namespace. When this occurs, containers can, for example, access each others' filesystem and memory,
as long as they share user and group IDs. These effects could be unexpected, especially if security (e.g. egress controls)
occurs in a sidecar container.
The No Exec policy prevents users from execing into running pods unless they have an exemption. This policy is typically enforced within a production environment, but run in report-only mode in dev and staging environments to facilitate debugging.
Execing into a pod can enable someone to do many nefarious things to that workload. Eventually this policy will also apply a taint label to the Pod to indicate that it should no longer be trusted and can be evicted.
Host bind mounts (also called hostPath
mounts) can be used to exfiltrate data from or escalate privileges on the host system. Using host bind mounts can cause unreliability of the node if it causes a partition to fill up.
The Docker socket bind mount provides API access to the host Docker daemon, which can be used for privilege escalation or otherwise control the container host. Using Docker sock mounts can cause unreliability of the node because of the extra workloads that the Kubernetes schedulers are not aware of.
Note: It is recommended to use the No Bind Mounts
policy to disable all hostPath
mounts rather than only this policy.
By default, an emptyDir
lacks a sizeLimit
parameter, and is disk-based;
a Pod with access to said emptyDir
can consume the Node's entire disk (i.e. the limit is unbounded) until the offending Pod is deleted or evicted, which can constitute a denial-of-service condition at the affected Node (i.e. DiskPressure).
This policy
- sets the configured default size when none is set for an
emptyDir
volume - reports a violation when the size is greater then the configured max size
policy_config:
mutate_empty_dir_size_limit:
maximum_size_limit: "1Gi"
default_size_limit: "512Mi"
Sets a default seccomp profile (runtime/default
or a configured one) for Pods if they have no existing seccomp configuration. The default seccomp policy for Docker and Containerd both block over 40 syscalls, many of which are potentially dangerous. The default policies are usually very compatible with applications, too.
The Mutate Default Seccomp Profile policy can be configured in the k-rail configuration file.
Example
policy_config:
policy_default_seccomp_policy: "runtime/default"
Docker image tags in a registry are mutable, so if you reference a tag without specifying the image digest someone or something could change the image you were using without you knowing.
You can obtain the immutable reference for an image with this command:
$ docker inspect --format='{{index .RepoDigests 0}}' alpine:3.8
alpine@sha256:dad671370a148e9d9573e3e10a9f8cc26ce937bea78f3da80b570c2442364406
You can also add the tag back in for it to be more human readable:
alpine:3.8@sha256:dad671370a148e9d9573e3e10a9f8cc26ce937bea78f3da80b570c2442364406
Host networking enables packet capture of host network interfaces and a bypass to some cloud meta data APIs, such as the GKE metadata API. The metadata API can be used to escalate access.
The host PID namespace can be used to inspect process environment variables (which often contain secrets). It can also potentially be used to dump process memory, modify kernel parameters, and insert kprobes+uprobes into the kernel to exfiltrate information.
Kernel Capabilities can be used to escalate to level of kernel API access available to the process. Some can enable loading kernel modules, changing namespace, load eBPF byte code in the kernel and other potentially dangerous things.
Privileged containers have all capabilities and also removes cgroup resource accounting.
Helm Tiller installations often have an unauthenticated API open to the cluster which provides a privilege escalation route to ClusterAdmin or NamespaceEditor.
Note: This policy only blocks images that /tiller
in their name from being deployed. It is not a robust policy and serves more as a reminder for engineers to seek an alternate route of deployment, such as using helm template
or isopod.
There are many malicious, poorly configured, and outdated and vulnerable images available in public Docker image repositories. Images must be sourced from configured trusted internal repositories or from an official Docker Hub repository.
The Trusted Image Repository policy can be configured in the k-rail configuration file.
Example
policy_config:
policy_trusted_repository_regexes:
- '^gcr.io/some-gcr-repo/.*' # private GCR repo
- '^k8s.gcr.io/.*' # official k8s GCR repo
- '^[A-Za-z0-9\-:@]+$' # official docker hub images
DEPRECATED - See Mutate Safe to Evict
below
The Kubernetes autoscaler will not evict pods using hostPath or emptyDir mounts unless they have this annotation:
cluster-autoscaler.kubernetes.io/safe-to-evict=true
This policy validates that Pods have this annotation. You'll probably find the mutation policy below more useful.
The Kubernetes autoscaler will not evict pods using hostPath or emptyDir mounts unless they have this annotation:
cluster-autoscaler.kubernetes.io/safe-to-evict=true
This policy mutates Pods that do not have the annotation specfied to be true
. It will not override existing annotations with false
.
You can also set the annoation on existing Pods with this one-liner:
$ kubectl get po --all-namespaces -o json | jq -r '.items[] | select(.spec.volumes[].hostPath or .spec.volumes[].emptyDir) | [ .metadata.namespace, .metadata.name ] | @tsv' | while IFS=$'\t' read -r namespace pod; do echo "\n NAMESPACE: $namespace \n POD: $pod \n"; kubectl annotate pod -n $namespace $pod "cluster-autoscaler.kubernetes.io/safe-to-evict=true"; done
There are cerntain images which require the enforcement of the ImagePullPolicy according to different user scenarios
- IfNotPresent It can reduce the unnecessary traffic (Auth and Download requests) to Image repository and reuse the image which is cached on the node
- Always It can be useful when it requires the absolute isolation in multi-tenant cluster, which prevents others to reuse the image cached on the node, for example: The image protected with ImagePullSecret from private repository is cached on the node after first successful pull, other user can directly pull from node without proper auth. However if we force the imagePullPolicy to Always, it would fail without proper ImagePullSecret
The Mutate Image Pull Policy can be configured in the k-rail configuration file.
Example
policy_config:
mutate_image_pull_policy:
IfNotPresent:
- '^gcr.io/repo/image1.*'
- '^gcr.io/repo/image2.*'
Always:
- '^gcr.io/private-repo/secretimage.*'
The Require Ingress Exemption policy requires the configured ingress classes to have an a Policy exemption to be used. This is typically useful if you want to gate the usage of public ingress.
The Require Ingress Exemption policy can be configured in the k-rail configuration file.
Example
policy_config:
policy_require_ingress_exemption_classes:
- nginx-public
Unique Ingress Host policy requires the configured ingress hosts to be unique across cluster namespaces. This is helps to prevent ingress host collisions.
For the Helm deployment, all configuration is contained in deploy/helm/values.yaml
.
By default, k-rail will "fail close" if it cannot be reached by the API server. k-rail can be changed to
"fail open" by changing the failurePolicy
directive from Fail
to Ignore
, in deploy/helm/values.yaml
.
See the Kubernetes docs for
more details.
In Kubernetes 1.15 and beyond, mutating admission webhooks (e.g. k-rail) can elect to be polled again, if a subsequent admission plugin
(such as another webhook) modifies an object the webhook has interacted with.
They do so with a
reinvocationPolicy
value of IfNeeded
; the Kubernetes default value is Never
, which does not reinvoke
the mutating admission webhook(s).
Since this is a newer type field, k-rail omits by default,
but operators can set a chosen value by commenting out reinvocationPolicy
in deploy/helm/values.yaml
.
See the associated KEP for more details on reinvocationPolicy
and admission plugin ordering.
Log levels can be set in the k-rail configuration file. Acceptable values are debug
, warn
, and info
. The default log level is info
.
All reporting and enforcement operations are logged in a structured json blob per event. It is useful to run policies in report-only mode, analyze your state in with the structured logs, and flip on enforcement mode when appropriate.
When global_report_only_mode
is toggled in the config, ALL policies run in report_only
mode, even if configured otherwise.
This mode must be false to have any policies in enforcement mode.
Policies can be enabled/disabled, and run in report-only or enforcement mode as specified in the config.
A folder to load policy exemptions from can be specified from config. Load exemptions by specifying the -exemptions-path-glob
parameter, and specify a path glob that includes the exemptions, such as /config/policies/*.yaml
.
For the Helm deployment, all policy and exemption configuration is contained in deploy/helm/values.yaml
.
The format of an exemption config is YAML, and looks like this:
---
# exempt all kube-system pods since they are largely provided by GKE
- resource_name: "*"
namespace: "kube-system"
exempt_policies:
- "*"
# malicious-pod needs host network to escalate access via GCE metadata API
- resource_name: malicious-pod
namespace: malicious
exempt_policies: ["pod_no_host_network"]
# allow everything
# - resource_name: "*"
# namespace: "*"
# username: "*"
# group: "*"
# exempt_policies: ["*"]
Some policies are configurable. Policy configuration is contained in the k-rail configuration file, and documentation for a policy's configuration can be found in the Supported policies heading above.
For the Helm deployment, all policy and exemption configuration is contained in deploy/helm/values.yaml
.
Policies must satisfy this interface:
// Policy specifies how a Policy is implemented
// Returns an optional slice of violations and an optional slice of patch operations if mutation is desired.
type Policy interface {
Name() string
Validate(ctx context.Context,
config policies.Config,
ar *admissionv1beta1.AdmissionRequest,
) ([]policies.ResourceViolation, []policies.PatchOperation)
}
Name()
must return a string that matches a policy name that is provided in configuration.
Validate accepts an AdmissionRequest, and the resource of interest must be extracted from it. See resource/pod.go
for an example of extracting PodSpecs from an AdmissionRequest. If mutation on a resource is desired, you can return a slice of JSONPatch operations and nil
for the violations.
Policies can be registered in internal/policies.go
. Any policies that are registered but do not have configuration provided get enabled in report-only mode.
If you see timeout events on resources, this may be because the k-rail
service is unreachable from the Kubernetes apiserver.
Newer versions (1.14+) of Kubernetes are not likely to have this issue if the MutatingWebhookConfiguration
failurePolicy
is set to Ignore
and timeoutSeconds
is set to a lower number (such as 5
or less).
To determine if this is occuring because the service is unreachable, check the kube-apiserver
logs. You will see logs similar to,
E0911 04:57:22.686526 1 dispatcher.go:68] failed calling webhook "k-rail.cruise-automation.github.com": Post https://k-rail.k-rail.svc:443/?timeout=30s: dial tcp 10.110.63.191:443: connect: connection refused
Checking kube-apiserver
logs depends on what Kubernetes distribution you use.
- For minikube and other self hosted (meaning Kubernetes runs its infra on itself) clusters, you can typically just view the logs for the
apiserver
pod in thekube-system
namespace. - For some non self-hosted clusters, such as GKE, you can download the
apiserver
logs through the Kubernetes proxy:
kubectl proxy --port 8001 &
curl http://localhost:8001/logs/kube-apiserver.log > /tmp/out.log
The apiserver tracks latency and status code metrics for webhooks. This may be useful for debugging timeouts or assurance of performance.
kubectl proxy --port 8001 &
curl -s http://localhost:8001/metrics | grep k-rail
This may be caused by the k-rail
service being unreachable. To determine this, see Resources are having timeout events.
If you need to make an exemption to a policy, see Policy exemptions.
By default, a 10 year certificate is generated during each apply of k-rail. Re-applying will renew it.
You can also check the expiration with this command:
$ kubectl get secret --namespace k-rail k-rail-cert -o json | jq -r '.data["cert.pem"]' | base64 -d | openssl x509 -noout -text | grep -A 3 "Validity"
Validity
Not Before: Oct 24 05:40:16 2019 GMT
Not After : Oct 21 05:40:16 2029 GMT
Subject: CN = k-rail.k-rail.svc
Copyright (c) 2019-present, Cruise LLC
This source code is licensed under the Apache License, Version 2.0, found in the LICENSE file in the root directory of this source tree. You may not use this file except in compliance with the License.