Skip to content
Hands-on IT IT Drills

Day 9: Authentication, RBAC and Admission

Goal: You give your application exactly the API permissions it needs and understand why the API server rejects a pod despite correct permissions. · Duration: approx. 75 min · Prerequisite: Day 8

Briefing

CKAD: “Understand authentication, authorization and admission control” (Application Environment, Configuration and Security, 25 %).

  Request --> AuthN ---------> AuthZ ------------> Admission -------------> etcd
              Who are you?     May you do that?    Is the object allowed?
              certificate,     RBAC: Role +        mutating, then validating:
              SA token         RoleBinding         PodSecurity, VAP, webhooks
              401              403                 403 / invalid
  • A pod talks to the API as a ServiceAccount: system:serviceaccount:<ns>:<name>.
  • RBAC only knows permissions, no denials. Without a matching binding: no.
mkdir -p ~/ckad/tag09 && cd ~/ckad/tag09
k create ns tag09
k config set-context --current --namespace=tag09

Drill 1: Who am I?

k auth whoami

Expected (shortened):

ATTRIBUTE   VALUE
Username    kubernetes-admin
Groups      [kubeadm:cluster-admins system:authenticated]

You are logged in with a client certificate and are a member of a group with full permissions. Your application gets its own ServiceAccount – and initially nothing at all:

k create sa notes-app
k auth can-i list pods --as=system:serviceaccount:tag09:notes-app

Expected:

serviceaccount/notes-app created
no

Save your progress? Sign up for free

Drill 2: Role and RoleBinding

k create role pod-reader --verb=get,list,watch --resource=pods
k create rolebinding notes-app-pod-reader --role=pod-reader \
  --serviceaccount=tag09:notes-app
k auth can-i list pods --as=system:serviceaccount:tag09:notes-app
k auth can-i delete pods --as=system:serviceaccount:tag09:notes-app
k auth can-i list pods -n default --as=system:serviceaccount:tag09:notes-app

Expected: yes, no, no – reading yes, deleting no, and only in its own namespace.

All permissions at once:

k auth can-i --list --as=system:serviceaccount:tag09:notes-app | grep -v '^ '

Expected (shortened):

Resources                                       Non-Resource URLs   Resource Names   ...
selfsubjectreviews.authentication.k8s.io        []                  []               ...
...
pods                                            []                  []               ...

The lines with /api, /healthz etc. (hidden here with grep -v) are allowed for every authenticated client. On this cluster, Calico’s own policy types show up as well.

Remember the notation: with --serviceaccount it is namespace:name, with --as system:serviceaccount:namespace:name.

Save your progress? Sign up for free

Drill 3: Use built-in ClusterRoles in a namespace

Kubernetes ships with view, edit and admin. Via a RoleBinding they apply only in the binding’s namespace:

k create sa ci-bot
k create rolebinding ci-bot-edit --clusterrole=edit --serviceaccount=tag09:ci-bot
k auth can-i create deployments.apps --as=system:serviceaccount:tag09:ci-bot
k auth can-i get secrets --as=system:serviceaccount:tag09:ci-bot
k auth can-i create rolebindings --as=system:serviceaccount:tag09:ci-bot
k auth can-i create deployments -n default --as=system:serviceaccount:tag09:ci-bot

Expected: yes, yes, no, no.

k create rolebinding notes-app-view --clusterrole=view --serviceaccount=tag09:notes-app
k auth can-i get configmaps --as=system:serviceaccount:tag09:notes-app
k auth can-i get secrets --as=system:serviceaccount:tag09:notes-app
k get rolebinding -o wide

Expected:

rolebinding.rbac.authorization.k8s.io/notes-app-view created
yes
no
NAME                   ROLE               AGE   USERS   GROUPS   SERVICEACCOUNTS
ci-bot-edit            ClusterRole/edit   7s                     tag09/ci-bot
notes-app-pod-reader   Role/pod-reader    15s                    tag09/notes-app
notes-app-view         ClusterRole/view   1s                     tag09/notes-app

Why? edit may read Secrets, view may not. Whoever grants edit also hands out the passwords in the namespace.

Save your progress? Sign up for free

Drill 4: A pod queries the API – 403, then 200

Every pod gets the token of its ServiceAccount mounted under /var/run/secrets/kubernetes.io/serviceaccount.

k create sa lister
cat > api-client.yaml <<'EOF'
apiVersion: v1
kind: Pod
metadata:
  name: api-client
spec:
  serviceAccountName: lister
  containers:
  - name: curl
    image: curlimages/curl:8.17.0
    command: ["sleep", "3600"]
EOF
k apply -f api-client.yaml
k wait --for=condition=Ready pod/api-client --timeout=90s
k exec api-client -- ls /var/run/secrets/kubernetes.io/serviceaccount

Expected:

serviceaccount/lister created
pod/api-client created
pod/api-client condition met
ca.crt
namespace
token

Query the ConfigMaps of its own namespace with token and CA:

k exec api-client -- sh -c 'D=/var/run/secrets/kubernetes.io/serviceaccount
curl -s --cacert $D/ca.crt -H "Authorization: Bearer $(cat $D/token)" \
  https://kubernetes.default.svc/api/v1/namespaces/tag09/configmaps' \
  | grep -E '"(kind|message|code)"'

Expected:

  "kind": "Status",
  "message": "configmaps is forbidden: User \"system:serviceaccount:tag09:lister\"
cannot list resource \"configmaps\" in API group \"\" in the namespace \"tag09\"",
    "kind": "configmaps"
  "code": 403

Authenticated yes (the name is in the message), authorized no. Now bind the role and repeat the same command:

k create rolebinding lister-view --clusterrole=view --serviceaccount=tag09:lister
k exec api-client -- sh -c 'D=/var/run/secrets/kubernetes.io/serviceaccount
curl -s -o /dev/null -w "%{http_code}\n" --cacert $D/ca.crt \
  -H "Authorization: Bearer $(cat $D/token)" \
  https://kubernetes.default.svc/api/v1/namespaces/tag09/configmaps'

Expected:

rolebinding.rbac.authorization.k8s.io/lister-view created
200

RBAC takes effect immediately – no pod restart needed.

Save your progress? Sign up for free

Drill 5: Admission – ValidatingAdmissionPolicy

RBAC asks “may you create Deployments?”, not “with how many replicas?”. That is checked by admission. A ValidatingAdmissionPolicy (VAP) is a CEL rule directly in the API server; the binding defines where it applies – here only in namespaces with a label:

cat > tag09-max-replicas.yaml <<'EOF'
apiVersion: admissionregistration.k8s.io/v1
kind: ValidatingAdmissionPolicy
metadata:
  name: tag09-max-replicas
spec:
  failurePolicy: Fail
  matchConstraints:
    resourceRules:
    - apiGroups: ["apps"]
      apiVersions: ["v1"]
      operations: ["CREATE", "UPDATE"]
      resources: ["deployments"]
  validations:
  - expression: "object.spec.replicas <= 3"
    message: "Höchstens 3 Replicas erlaubt"
---
apiVersion: admissionregistration.k8s.io/v1
kind: ValidatingAdmissionPolicyBinding
metadata:
  name: tag09-max-replicas
spec:
  policyName: tag09-max-replicas
  validationActions: ["Deny"]
  matchResources:
    namespaceSelector:
      matchLabels:
        tag09-limits: "on"
EOF
k apply -f tag09-max-replicas.yaml
k create deploy big --image=nginx:1.29 --replicas=5 --dry-run=server

Expected:

validatingadmissionpolicy.admissionregistration.k8s.io/tag09-max-replicas created
validatingadmissionpolicybinding.admissionregistration.k8s.io/tag09-max-replicas created
deployment.apps/big created (server dry run)

The label is still missing. --dry-run=server runs through admission but stores nothing.

k label ns tag09 tag09-limits=on

The API server needs one or two seconds to pick up new policies and labels – if you type faster, you still slip through.

k create deploy big --image=nginx:1.29 --replicas=5
k create deploy small --image=nginx:1.29 --replicas=2

Expected:

error: failed to create deployment: deployments.apps "big" is forbidden:
ValidatingAdmissionPolicy 'tag09-max-replicas' with binding 'tag09-max-replicas' denied
request: Höchstens 3 Replicas erlaubt
deployment.apps/small created

Save your progress? Sign up for free

Drill 6: Admission – Pod Security

The built-in PodSecurity plugin checks pods against the levels privileged, baseline or restricted. It is switched on with a namespace label. First test what would happen to the running pods:

k label ns tag09 pod-security.kubernetes.io/enforce=restricted --dry-run=server

Expected (shortened):

Warning: existing pods in namespace "tag09" violate the new PodSecurity enforce level
"restricted:latest"
Warning: api-client (and 2 other pods): allowPrivilegeEscalation != false, ...
namespace/tag09 labeled (server dry run)

The two others are the nginx pods of small. Running pods are left untouched – the check only happens at creation:

k label ns tag09 pod-security.kubernetes.io/enforce=restricted
k run notes --image=172.16.0.126:30500/notes-api:1.0 --env=NOTES_API_KEY=lab-key-123

Expected (shortened, wrapped):

Warning: existing pods in namespace "tag09" violate the new PodSecurity enforce ...
namespace/tag09 labeled
Error from server (Forbidden): pods "notes" is forbidden: violates PodSecurity
"restricted:latest": allowPrivilegeEscalation != false (container "notes" must set
securityContext.allowPrivilegeEscalation=false), unrestricted capabilities
(container "notes" must set securityContext.capabilities.drop=["ALL"]), runAsNonRoot !=
true (pod or container "notes" must set securityContext.runAsNonRoot=true),
seccompProfile (pod or container "notes" must set securityContext.seccompProfile.type
to "RuntimeDefault" or "Localhost")

The message is the repair manual. lab-key-123 is purely a lab value.

cat > notes.yaml <<'EOF'
apiVersion: v1
kind: Pod
metadata:
  name: notes
spec:
  securityContext:
    runAsNonRoot: true
    seccompProfile:
      type: RuntimeDefault
  containers:
  - name: notes
    image: 172.16.0.126:30500/notes-api:1.0
    env:
    - name: NOTES_API_KEY
      value: lab-key-123
    securityContext:
      allowPrivilegeEscalation: false
      capabilities:
        drop: ["ALL"]
EOF
k apply -f notes.yaml
k wait --for=condition=Ready pod/notes --timeout=60s

Expected:

pod/notes created
pod/notes condition met

runAsNonRoot works because the image runs as UID 65532. nginx (runs as root) would additionally need runAsUser and a different image (nginxinc/nginx-unprivileged) here.

Warning: For a Deployment you only get a Warning; the Deployment is created – but its pods are not. The reason is shown by k get events --field-selector reason=FailedCreate.

Save your progress? Sign up for free

Drill 7: Which admission plugins are running?

Read only, change nothing – on cp1:

sudo grep -E 'authorization-mode|admission' \
  /etc/kubernetes/manifests/kube-apiserver.yaml
k exec -n kube-system kube-apiserver-cp1 -- kube-apiserver -h \
  | grep -o 'in addition to default enabled ones ([^)]*)'

Expected (shortened):

    - --authorization-mode=Node,RBAC
    - --enable-admission-plugins=NodeRestriction
in addition to default enabled ones (NamespaceLifecycle, LimitRanger, ...
PodSecurity, ... MutatingAdmissionWebhook, ValidatingAdmissionPolicy,
ValidatingAdmissionWebhook, ResourceQuota)

So PodSecurity, ValidatingAdmissionPolicy and ResourceQuota (Day 7) are active without any extra work; kubeadm only additionally enables NodeRestriction.

Save your progress? Sign up for free

Exercise: Speed Drills

  1. The (new) ServiceAccount reporter should be allowed to read Deployments in tag09 (get, list). Target time: 60 s
  2. May ci-bot delete pods in tag09? Target time: 15 s
  3. Additionally set a warning for violations of baseline on tag09. Target time: 20 s

Solution

k create sa reporter
k create role deploy-reader --verb=get,list --resource=deployments.apps
k create rolebinding reporter-deploy-reader --role=deploy-reader \
  --serviceaccount=tag09:reporter
k auth can-i list deployments.apps --as=system:serviceaccount:tag09:reporter
k auth can-i delete pods --as=system:serviceaccount:tag09:ci-bot
k label ns tag09 pod-security.kubernetes.io/warn=baseline

Expected:

serviceaccount/reporter created
role.rbac.authorization.k8s.io/deploy-reader created
rolebinding.rbac.authorization.k8s.io/reporter-deploy-reader created
yes
yes
namespace/tag09 labeled

Save your progress? Sign up for free

Checkpoint

k auth can-i get secrets --as=system:serviceaccount:tag09:notes-app

Expected: no

k get ns tag09 -L pod-security.kubernetes.io/enforce,tag09-limits

Expected:

NAME    STATUS   AGE   ENFORCE      TAG09-LIMITS
tag09   Active   5m    restricted   on
k get pod notes -o jsonpath='{.status.phase}{"\n"}'

Expected: Running

Challenge

Scale small to 5 replicas with k scale. The policy does not kick in – find the reason and close the gap.

Hint: k scale does not change the Deployment but its scale subresource.

Solution

k scale deploy small --replicas=5
k scale deploy small --replicas=2
sed -i 's|\["deployments"\]|["deployments", "deployments/scale"]|' \
  tag09-max-replicas.yaml
k apply -f tag09-max-replicas.yaml
k get validatingadmissionpolicy tag09-max-replicas \
  -o jsonpath='{.spec.matchConstraints.resourceRules[0].resources}{"\n"}'
k scale deploy small --replicas=5

Expected:

deployment.apps/small scaled
deployment.apps/small scaled
validatingadmissionpolicy.admissionregistration.k8s.io/tag09-max-replicas configured
validatingadmissionpolicybinding.admissionregistration.k8s.io/tag09-max-replicas ...
["deployments","deployments/scale"]
The deployments "small" is invalid: : ValidatingAdmissionPolicy 'tag09-max-replicas'
with binding 'tag09-max-replicas' denied request: Höchstens 3 Replicas erlaubt

The expression matches both: the Scale object also has spec.replicas.

Clean up – the policy is cluster-wide:

k config set-context --current --namespace=default
k delete ns tag09
k delete validatingadmissionpolicybinding tag09-max-replicas
k delete validatingadmissionpolicy tag09-max-replicas

Save your progress? Sign up for free

Exam Tip

  • After every RBAC task, prove that it works with k auth can-i … --as=system:serviceaccount:<ns>:<sa>. Specify resources outside the core group with their group: deployments.apps.
  • If PodSecurity rejects a pod, transfer the message point by point into the securityContext. For Deployments, the errors are in the ReplicaSet’s events.
  • Docs: kubernetes.io → “Pod Security Standards” contains the ready-made fields for restricted.

Enterprise Note

  • Humans log in via OIDC (Keycloak, Entra ID); bind permissions to groups, never to individuals.
  • automountServiceAccountToken: false everywhere an application does not need the API.
  • Set Pod Security via labels on all namespaces, first with warn and audit, then enforce.
  • Complex rules (image provenance, mandatory labels) are handled by Kyverno, OPA Gatekeeper or VAPs with parameters; version policies in Git.