BRev Port 6443 — Kubernetes API server
Port 6443 is the kube-apiserver: the single front door to a Kubernetes cluster, and the only component that talks to etcd. Every kubectl command, every controller, every CI deployment and every node registration goes through it — which is why, unusually for a control-plane service, exposing it is a legitimate design rather than a mistake. Managed clusters on EKS, GKE and AKS publish it by default, because that is how you administer them.
What makes that acceptable is the authentication model: TLS with client certificates, service-account tokens or an OIDC provider, followed by RBAC authorisation on every request. Our reference marks it care, and the care is a short list. Anonymous authentication must be off — the historical default allowed unauthenticated requests to reach the API, and a permissive RBAC binding on system:anonymous or system:unauthenticated turns that into cluster-admin. Authorized-network ranges should be set so the API answers only your addresses. And the certificate authority for the cluster is what your clients pin, so it must be handled carefully.
Two neighbours matter in a scan. etcd on 2379/2380 holds every secret in the cluster in plain form and must never be reachable outside the control plane — an exposed etcd is a complete compromise with no authentication step. The kubelet on 10250 is similar: reachable and misconfigured, it executes commands in pods. 6443 is the port meant to be public; those two are not.
Port 6443 at a glance
| Field | Value | Detail |
|---|---|---|
| Port | 6443 | Registered |
| Service | Kubernetes API | The kube-apiserver endpoint. (TCP) |
| Category | Infrastructure | the plumbing: DNS, directory, routing, monitoring |
| Exposure | Care | Expose with care |
Should port 6443 face the internet? Managed clusters expose it by design — restrict authorised networks, never leave anonymous auth on.
Legitimately public in some setups, but only hardened — current version, real authentication, rate limits, and ideally an allow-list.
What commonly goes wrong
“Unable to connect to the server: x509: certificate signed by unknown authority”
The kubeconfig does not carry the cluster’s CA, or the API server is presenting a certificate for a name you are not using. Fix the CA data in the kubeconfig rather than reaching for --insecure-skip-tls-verify, which removes exactly the protection that makes exposing 6443 reasonable.
kubectl works from one network and times out from another
Authorized networks (or the equivalent security group) restricting which source addresses reach the endpoint — which is the control working. Add the range deliberately rather than opening it to the world; for CI, a fixed egress address or a private endpoint is the durable answer.
Check port 6443 yourself
No browser can open a raw TCP socket, so no web page can honestly tell you whether this port is open on a remote host. These are the real commands, generated for port 6443 — swap example.com for the host you are actually testing.
Is port 6443 reachable on another host?
nc -vz example.com 6443"succeeded!" means the TCP handshake completed — the port is open. "Connection refused" is closed; a hang until timeout means filtered (a firewall dropped it silently).
nmap -Pn -p 6443 example.comThe only common tool that distinguishes open / closed / filtered. -Pn skips the ping sweep, so it still works against hosts that ignore ICMP.
Test-NetConnection example.com -Port 6443Read TcpTestSucceeded: True = open. It is built in — nothing to install.
telnet example.com 6443A blank screen or a banner means it connected. "Connection refused" means closed. Quit with Ctrl-] then `quit`.
curl -v --max-time 5 telnet://example.com:6443Handy when telnet is not installed: "Connected to …" proves the TCP handshake, without curl trying to speak HTTP.
What on this machine is holding port 6443?
sudo ss -tlnp | grep :6443Lists the listening socket and the PID/program holding it. Drop sudo and you get the socket but not the process name.
sudo lsof -i :6443The macOS answer to "address already in use". Kill it with `kill -9 <PID>` once you are sure what it is.
netstat -ano | findstr :6443The last column is the PID. Match it in Task Manager’s Details tab, or `taskkill /PID <pid> /F`.
No web page can tell you whether a port is open on a remote host
A browser cannot open a raw TCP socket — there is no API for it, by design. So a site that reports “port 3389 is OPEN on 203.0.113.10” is not checking from your browser; it is asking its own server to scan on your behalf. That works, but it means their server sees the host you asked about, scans from their address rather than yours, and answers a question about their network path, not yours.
We don't have a server here and won't pretend to. So this page does three honest things instead:
| Port | Proto | Service | What it is | Exposure | |
|---|---|---|---|---|---|
| 6443 | tcp | Kubernetes API | The kube-apiserver endpoint.Managed clusters expose it by design — restrict authorised networks, never leave anonymous auth on. | Care |
http://host:port/ and times it. A connection that is established proves a socket is open; an instant refusal usually means nothing is listening; anything else is honestly reported as “can't tell”. Each attempt gives up after 2.5s.nc -vz example.com 443
nmap -Pn -p 443 example.com
telnet example.com 443
curl -v --max-time 5 telnet://example.com:443
This page cannot answer that, and neither can any other page — from your browser. Opening a TCP connection to an arbitrary host and port is not something JavaScript is allowed to do. Every “open port checker” that gives you a green tick is running the scan on their server. Which is fine, as long as you know what you are agreeing to: they learn which host and port you were worried about, and the answer they give you is about the route from their data centre — not from the internet as your users see it.
What to do instead
Run the check from outside your own network. Copy an nmap or nc line from the card above and run it from a machine that is not behind the same router — a phone on mobile data with a terminal app, a cheap VPS, a colleague's laptop. Running it from inside the LAN tests the LAN, not the internet, and will happily say “open” while the world sees nothing.
Then check the thing that is actually blocking it, in this order: the service is bound to0.0.0.0 and not 127.0.0.1 · the host firewall (ufw, firewalld, Windows Defender Firewall) · the cloud security group or the router's port-forward rule · and finally your ISP, which commonly blocks 25, 80 and 445 on residential lines whatever you configure.
What a browser genuinely can verify
One thing: that an HTTPS service answers. If what you actually want to know is “is my site up and how fast is it”, the latency test measures real round trips to a CORS-open URL. If the name doesn't resolve, or resolves somewhere unexpected, that is a DNS problem, not a port problem — and it is the more common cause of “my port is closed” than an actual firewall.
An IP address finds the machine; a port picks which program on it gets the traffic. TCP sets up a connection first — a handshake, ordered delivery, retransmission — which is why the web, email and SSH use it. UDP just sends the packet with no handshake and no guarantee it arrives, which is why DNS, video calls and games use it: late data is worse than lost data.
That difference is why a UDP port is so much harder to test. TCP either completes a handshake or it doesn't, so a tool can tell you. A UDP service that has nothing to say to an unexpected packet stays silent — and silence looks identical to a firewall dropping it.
The three ranges
0–1023Assigned by IANA to core services. On Unix, binding one needs root (or CAP_NET_BIND_SERVICE).1024–49151Registered by vendors for their applications. Any user can bind these — which is why every dev server lives here.49152–65535Never assigned. The OS hands these out as the SOURCE port of your outgoing connections. (Linux actually uses 32768–60999 by default.)Closed is not the same as filtered
Closed means the host answered: the packet arrived, nothing is listening, and you got an immediate refusal (a TCP RST). Filtered means nothing answered at all — a firewall dropped your packet in silence, so you wait for a timeout and learn nothing. The difference matters: closed proves the host is reachable and the port is free; filtered tells you only that something between you and it is discarding traffic.nmap reports the two separately. A browser cannot distinguish them, which is why this page reports a timeout as “can't tell” rather than guessing.
The one habit worth keeping
Bind services to 127.0.0.1 instead of 0.0.0.0 unless something outside the machine genuinely needs them, and reach them over SSH or a VPN. Every database on the never-expose list above ships with that advice, and the leaks all begin the same way: a default bind address, a cloud instance with a public IP, and a scanner that found it within the hour.
The blocked-port list this page uses is the WHATWG Fetch Standard's “bad port” table — the ports every browser refuses to connect to (Chrome calls it ERR_UNSAFE_PORT). Local probing is additionally gated by Chrome's Local Network Access permission, which replaced Private Network Access and ships in Chrome 142.
FAQ
Is it safe to expose the Kubernetes API on 6443?
It is the intended design, provided anonymous authentication is disabled, RBAC is not permissively bound, authorized networks are set, and audit logging is on. It is the only control-plane port that should ever be publicly reachable.
What about etcd on 2379?
Never expose it. etcd stores every Kubernetes object including secrets, and a reachable, unauthenticated etcd is a complete cluster compromise with no login to defeat. It belongs on the control-plane network only.
Why is my cluster on 443 instead of 6443?
Managed providers commonly front the API server with a load balancer on 443 so it passes through restrictive client networks. Same API, same authentication — self-managed clusters and kubeadm defaults use 6443 directly.