Skip to main content
USA-Based Digital Agency

AI security service · Updated

MCP server security audit

An MCP server gives AI assistants a path into your systems. We audit MCP servers, yours or third-party ones you plan to connect, against the Model Context Protocol specification’s security requirements: authorization, token handling, tool scopes, request validation, consent for local commands, and the dependencies the server runs on.

By George Shvaya, founder of Webvello · Published

Key takeaways

  • We check servers against the current MCP specification (version 2026-07-28) and its Security Best Practices page.
  • The audit covers confused deputy, token passthrough, SSRF, state handle hijacking and local server compromise, the risks the spec documents.
  • We review tool descriptions and scopes for tool poisoning exposure and excess permissions.
  • You get findings with spec references and fixes, and we can implement them.
  • We audit only servers you own or are authorized to assess.

What does an MCP security audit check?

Each area maps to a section of the MCP specification or a documented attack.
Areas covered by a Webvello MCP server security audit
AreaWhat we check
AuthorizationWhether remote access requires authorization; OAuth 2.1 flow, resource indicators and token audience validation
Token handlingNo token passthrough; downstream APIs use the server’s own scoped credentials
Proxy consentPer-client consent where the server proxies a third-party authorization server (confused deputy)
Request validationURLs the server or client fetches are restricted, and internal and metadata addresses are blocked (SSRF)
State and sessionsState handles are never treated as authentication
ToolsEach tool’s description, inputs, data access and side effects; read and write separation; poisoning exposure
Local serversConsent before commands run, sandboxing, and what the server can reach on the machine
DependenciesMCP SDKs, transport libraries and tooling pinned and free of known critical issues

The checklist follows the MCP Security Best Practices page. Background on each risk is in our MCP security risks guide.

How does the audit work?

Code review first, then testing against a running server.
  1. 1

    Scope

    Which servers, versions and environments, and whether we get source code. Authorization in writing.
  2. 2

    Code and configuration review

    Auth, token validation, tool definitions, input and output handling, logging and dependencies.
  3. 3

    Live testing

    Against a test deployment: authorization bypass attempts, token audience checks, URL validation and tool behavior with untrusted content.
  4. 4

    Report

    Findings with severity, evidence, the spec section they relate to, and a fix.
  5. 5

    Fix and retest

    We fix the issues or support your team, then confirm they are closed.

Tell us about your MCP server

What it connects to, who uses it and where it runs. We will reply with what an audit would cover.

Can you audit a third-party MCP server before we connect it?

Yes, within what its license and source availability allow.

Before connecting an open-source or vendor MCP server to an assistant that has access to your data, we review what each tool does, what it reads, what credentials it needs and how it is maintained, and recommend whether to connect it, connect it with limits, or skip it. For servers we build ourselves, see MCP server development.

Pages that go deeper on the pieces this one touches.

Common questions

Straight answers to what owners ask us first.

Which MCP specification version do you audit against?

The version your server implements. The current version is 2026-07-28; if your server targets an earlier one, we note where its security guidance differs and what upgrading would change.

Our MCP server only runs locally. Does it need an audit?

Local servers run with the user’s permissions on their machine, and the specification requires consent before executing commands. Developer tooling has had critical vulnerabilities, so local servers are worth reviewing too.

Do you also review the AI app that uses the server?

We can. Our AI application security review covers the host app and agent: prompt injection, permissions and output handling. Many teams combine the two.

Close the path before someone finds it

Book a call or send a short description, and we will scope an audit of your MCP server.