CVE-2026-46519: mcp-server-kubernetes: auth bypass enables full cluster RCE

## Summary `mcp-server-kubernetes` exposes three environment variables (`ALLOW_ONLY_READONLY_TOOLS`, `ALLOW_ONLY_NON_DESTRUCTIVE_TOOLS`, `ALLOWED_TOOLS`) documented as access controls for restricting which Kubernetes operations are available. These controls are enforced at the tool discovery layer (`tools/list`) but not at the execution layer (`tools/call`). Any client that knows a tool name can invoke it directly regardless of the configured restriction mode. The access control was effectively cosmetic. Fixed in v3.6.0. ## Impact An attacker or misconfigured AI agent with network access to the MCP server's HTTP endpoint could invoke any Kubernetes tool regardless of the restriction mode configured by the operator -- including `kubectl_delete`, `exec_in_pod`, `kubectl_generic`, and `node_management`. The project explicitly supports and documents multi-client HTTP deployment scenarios (Streamable HTTP and SSE transports, in-cluster deployments, Codex CLI and Gemini CLI integrations). In these deployments, operators relied on the tool restriction env vars to enforce least-privilege access across users or roles. The bypass invalidated that model entirely. Severity scales with the Kubernetes service account's permissions. In environments where the MCP server runs with `cluster-admin` (common in dev/staging), this is equivalent to full cluster compromise for any client that can reach the endpoint. The `MCP_AUTH_TOKEN` / `X-MCP-AUTH` mechanism controls who can reach the endpoint but provides no per-tool authorization. An authenticated client restricted to `ALLOWED_TOOLS=kubectl_get` could still invoke `kubectl_delete` after authentication. ## Root Cause In `src/index.ts`, the `ListToolsRequestSchema` handler applied the configured filtering logic before returning available tools. The `CallToolRequestSchema` handler dispatched directly by tool name with no equivalent check -- every tool was reachable unconditionally. ## Proof of Concept Tested across all three restriction modes against a live kind cluster. In each case, `kubectl_delete` was absent from `tools/list` but executed successfully via a direct `tools/call` request: ```shell curl -s http://<HOST>:3003/mcp \ -H 'Content-Type: application/json' \ -H 'Accept: application/json, text/event-stream' \ -d '{"jsonrpc":"2.0","id":2,"method":"tools/call","params":{"name":"kubectl_delete","arguments":{"resourceType":"pod","name":"test-pod","namespace":"default"}}}' ``` Result: `{"result":{"content":[{"type":"text","text":"pod \"test-pod\" deleted\n"}]}}` Confirmed across `ALLOW_ONLY_READONLY_TOOLS=true`, `ALLOW_ONLY_NON_DESTRUCTIVE_TOOLS=true`, and `ALLOWED_TOOLS=kubectl_get`. ## Remediation The fix applies the same filtering logic from `ListToolsRequestSchema` at the start of the `CallToolRequestSchema` handler, returning an error for any tool call outside the active allowed set. Fixed in v3.6.0. ## Credit Discovered by [Francisco Rosales](https://www.linkedin.com/in/francisco-rosales-celis/) of [Manifold Security](https://manifold.security), coordinated by [Ax Sharma](https://www.linkedin.com/in/axsharma/), Head of Research at Manifold Security.

An attacker operating a legitimately authenticated but restricted MCP client — for example, a developer AI assistant granted ALLOWED_TOOLS=kubectl_get — first calls tools/list and confirms they see only kubectl_get in the response. They then craft a direct JSON-RPC tools/call request targeting kubectl_delete with a target pod name and namespace, exactly as documented in the advisory's PoC. The server dispatches the call unconditionally, returning confirmation of pod deletion. In a staging cluster running with cluster-admin, the attacker iterates through exec_in_pod to access running containers, harvests secrets from pod environments, and pivots to the Kubernetes API server. In an agentic pipeline scenario, a compromised or prompt-injected AI agent with access to the MCP server endpoint exploits the same bypass to perform destructive Kubernetes operations far beyond its intended scope, potentially taking down production workloads or exfiltrating sensitive configuration data.