GHSA-m932-63jc-4jqqHigh

k8s-mcp-server: Missing authentication on default SSE/HTTP transport allows unauthenticated Kubernetes tool execution

Published
November 19, 2026
Last Modified
November 19, 2026

🔗 CVE IDs covered (1)

📋 Description

Summary

When k8s-mcp-server runs in its default SSE transport mode, the MCP tools/call endpoint is dispatched with no authentication, session token, or Origin check. Any client that can reach the listening port can enumerate and invoke the server's Kubernetes tools (including destructive/write tools) without credentials, executing with the server's configured cluster privileges.

Affected

Commit be87a6dd72bb6ec1d4886b8548843b99a4638b12 (Merge PR #130) MCP library: github.com/mark3labs/mcp-go v0.57.0 Default mode: sse (SERVER_MODE default), default port 8080, bind 0.0.0.0

Root cause (static trace: handleSSE -> handleMessage -> tools/call)

  • main.go L129: server.NewSSEServer(s) is created with no auth option and no contextFunc; L130 sse.Start(":"+port) binds all interfaces (0.0.0.0).
  • mcp-go sse.go handleSSE (GET /sse): no auth header check; the server mints a sessionId for any caller. The sessionId is a correlation value, not a credential.
  • handleMessage (POST /message): the only gates are method==POST and a loadable sessionId. No Authorization/Bearer/token/Origin/CSRF check before s.server.HandleMessage(ctx, rawMessage) dispatches to the tools/call handler.
  • Only Access-Control-Allow-Origin: * is present (CORS relaxation, not auth). Write tools (Create/Delete/RolloutRestart, Helm Install/Upgrade/Uninstall/Rollback) are all registered unless -read-only is set (default readOnly=false).

Runtime reproduction (no auth headers on any request)

Started: k8s-mcp-server -mode sse -port 8080 with a dummy non-routable KUBERNETES_SERVER=https://127.0.0.1:1 (no real cluster touched).

GET /sse -> HTTP 200; server mints /message?sessionId=93b3... POST /message initialize -> HTTP 202; result returned over SSE POST /message tools/list -> HTTP 202; FULL tool list returned (incl. destructive) POST /message tools/call getAPIResources -> HTTP 202; handler EXECUTED: error -32603 "... Get https://127.0.0.1:1/api...: dial tcp 127.0.0.1:1: connect"

The tools/call response is NOT HTTP 401/403 but an in-handler cluster-connection error, proving the RPC passed any authorization gate and reached the getAPIResources handler. With a valid kubeconfig, the same path returns real cluster data. ss -ltnp confirms LISTEN *:8080 (all interfaces, externally exposed).

Impact

An unauthenticated attacker with network reach to the port can enumerate tools and invoke Kubernetes operations with the server's privileges — cluster read and, with default readOnly=false, state-changing operations (Create/Delete/Helm). CWE-306 (Missing Authentication for Critical Function).

Why this is a defect, not by-design

Default mode is HTTP (sse) on 0.0.0.0:8080, containerized deployment is the normal scenario, and there is no "trusted client only" / auth-required note in the docs. MCP HTTP transports are expected to be loopback-only or gated; this one is neither.

Suggested fix

  • Add an authorization gate before tools/call on the HTTP transports: e.g. WithSSEContextFunc to validate an Authorization bearer/token, or fail-closed when no token is configured in HTTP mode.
  • Default the bind to 127.0.0.1 and require explicit opt-in for external binding.
  • Document that HTTP mode must sit behind an authenticating proxy; recommend STDIO by default.

Reproduction was done against the actual built binary and the shipped SSE path (not an extract).

Credit

bebold6133

🎯 Affected products1

  • go/github.com/reza-gholizade/k8s-mcp-server:< 1.1.0

🔗 References (7)