CVE-2026-18673

CVE-2026-18673 is a trust-model failure, not merely an information leak. The kuma-dp data plane proxy exposes a TCP readiness service on port 9902 bound to all interfaces, bridging the internal Envoy admin API to the cluster network. In Kubernetes, pod networking is flat — if an attacker lands on any pod, they can reach this port on every other kuma-dp instance. That turns the cluster network into a trusted boundary it should never be. The exposed data is substantial. The readiness endpoint returns full Envoy configuration including cluster topology, listener definitions, upstream dependencies, and the mesh trust bundle. The trust bundle reveals the CA structure and mesh membership — an attacker can map authentication patterns before attempting to forge credentials. This transforms noisy exploitation into surgical operations. The CVSS 5.3 score reflects the immediate, bounded exposure but underweights the multi-stage attack path: reconnaissance enabling credential forging at scale. The 'destructive actions blocked' framing in the advisory creates false comfort. Today's implementation may block destructive operations, but the architectural problem — a TCP channel from untrusted network space to an admin API surface — persists. Future Envoy changes could expose more through this same channel. The fix likely restricts port 9902 to localhost or adds authentication, but this creates operational friction: kubelet health checks across nodes may fail, and operators may disable readiness probes entirely, trading a medium-severity disclosure for degraded reliability. Check your deployments now. Verify whether port 9902 is listening on non-localhost interfaces. If it is, assess whether your threat model assumes the cluster network is trusted — it shouldn't. Consider whether similar patterns exist in other service mesh components: admin interfaces, debug endpoints, and readiness probes consistently become trust boundary violations under operational pressure. The pattern is structural, not accidental.

Reviewed through automated stages and approved by a human before publication.