How to Evaluate an MCP Server Before You Connect It: A Security Checklist for Buyers
Before you connect an MCP server to an AI assistant, check three things: where it runs, which token it accepts and how much access it asks for. The Model Context Protocol project publishes a security best practices page that names the main attack classes and the controls that stop them. This guide turns that page, as read on 10 November 2026, into questions a buyer or a team lead can put to a vendor or an engineer. It does not replace a security review of your own environment.
Key facts
- The MCP security page names these attack classes: confused deputy, token passthrough, server-side request forgery, state handle hijacking, local server compromise, OAuth authorization URL attacks and mix-up attacks, plus scope design problems.
- Token passthrough is called an anti-pattern and is explicitly forbidden by the authorization specification. MCP servers must not accept any token that was not explicitly issued for the MCP server.
- Clients that support one-click local server setup must show the exact command, without truncation, and get explicit approval before running it.
- Local servers run with the same privileges as the client, so the page recommends sandboxing with minimal default privileges.
- Broad wildcard scopes are listed among common mistakes. The recommended model starts with a minimal scope set and raises access step by step.
Question 1: where does the server run?
A local server is a program that runs on the user's machine, often started by a command in a configuration file. The MCP page lists what can go wrong: a malicious startup command in a client configuration, a malicious payload inside the server itself, or an insecure server left running on localhost and reached through DNS rebinding. The result can be arbitrary code execution with the privileges of the client.
The page's controls are concrete. The client must show the full command and require approval. It should warn about commands that use sudo, delete files or reach the home directory, and ideally run servers in a sandbox with limited file and network access. A server meant to run locally should use the stdio transport, which limits access to the client, or require an authorization token for HTTP.
Ask: does this server run on laptops or on a managed host, who reviewed the startup command, and is it sandboxed?
Question 2: which token does it accept?
Token passthrough means a server accepts a token from the client without checking that it was issued to the server, then forwards it to a downstream API. The page says this breaks an OAuth security boundary and can bypass controls such as rate limiting, blur the audit trail and turn the server into a proxy for data theft if a token is stolen. Its rule is short: a server must not accept tokens that were not explicitly issued for it.
Ask: does the server validate the audience of every token, and can the vendor show where in the docs or code?
Question 3: how much access does it ask for?
The page warns that a token granted broad scopes, such as full access to files or databases, magnifies the damage if it leaks through logs or memory scraping. It lists publishing every possible scope, using wildcard scopes and bundling unrelated privileges among the common mistakes. The recommended design starts with a minimal set of low-risk read and discovery scopes and asks for more only when a privileged operation is first attempted.
Ask: what is the smallest scope set that makes the server useful, and does the vendor log each elevation?
Question 4: if it is a proxy, does it ask for consent per client?
The confused deputy attack hits MCP proxy servers that sit in front of a third-party API with one shared client ID. Under the conditions the page lists, a malicious client can obtain an authorization code without the user's consent. The fix is that the proxy must keep a registry of approved clients per user, show its own consent page naming the client and scopes, and match redirect addresses exactly.
Ask: does the server show its own consent screen before forwarding a user to the third-party login?
Question 5: what can it fetch on your network?
A malicious server can place internal addresses, such as cloud metadata endpoints or private IP ranges, in the URLs a client fetches during sign-in. That is a server-side request forgery risk. The page tells clients to require HTTPS, block private and reserved address ranges and consider an egress proxy, and it advises against writing your own IP validation because attackers use encoding tricks that custom parsers miss.
Ask: if the client runs on a server in your cloud, what outbound network rules limit it?
What to do
- Keep a register of every MCP server in use, with owner, host and scopes.
- Prefer remote servers from vendors that answer the five questions above in writing.
- Review the exact startup command of any local server before you approve it.
- Grant the smallest scope set first and add access when a task needs it.
- Re-check the vendor against the MCP page when the specification version changes.
Related coverage: the MCP SDK remote code execution issue and the widening authentication gap in MCP servers. Browse tools that offer MCP integrations in the AI tool directory.
Questions people ask
What is token passthrough in MCP?
It is when an MCP server accepts a token from a client without confirming it was issued to that server, then passes it on to another API. The MCP authorization specification forbids it.
Are local MCP servers safe?
They run with the same privileges as the client, so a malicious or buggy one can execute code or read files. The MCP page says clients must show the full command and get approval, and recommends sandboxing.
What is a confused deputy attack on MCP?
It lets a malicious client obtain an authorization code from a proxy server without the user's consent, using a shared client ID, dynamic client registration and a saved consent cookie. Per-client consent stops it.
Sources