
Somewhere in your organization right now, a developer is pasting a few lines of JSON into their coding assistant’s config file. That snippet starts an MCP server with a personal access token to your source code host. It works immediately. It never goes through security review. And it now gives an AI model the ability to read, write, and push code on that developer’s behalf.
Multiply that by every engineer, analyst, and power user experimenting with AI tools, and you have shadow MCP: servers that connect AI agents to business systems without anyone approving them.
This article explains where shadow MCP servers come from, the MCP security risks they create, how to find them, and how to bring them under control without shutting down the work people are trying to do.
What is a shadow MCP server?
A shadow MCP server is a Model Context Protocol server that’s connected to an AI client or agent without review, approval, or registration by the organization’s security or AI governance team.
MCP servers expose tools that let AI models take actions in other systems: query a database, update a ticket, send a message, run a command. When those servers are approved, they’re part of a governed integration. When nobody approved them, they’re shadow AI with a direct line into production systems.
Shadow MCP is harder to spot than other shadow AI because it doesn’t look like a new app. It looks like a configuration line, a package import, or an outbound API call from a tool your team already uses.
Where shadow MCP servers come from
Most shadow MCP servers are created by well-meaning people solving real problems. They usually show up in four places.
Developer laptops
Coding assistants and AI-enabled IDEs make it easy to add local MCP servers. A developer copies a config snippet from a README, adds a token, and the server starts with the IDE. Each one lives on a single machine, often with a long-lived personal credential.
AI assistants and desktop apps
Employees connect AI assistants to email, file storage, and SaaS tools through custom connectors. Some of those connectors point to MCP servers run by third parties the organization has never evaluated.
Code repositories
Product teams build agents that call MCP servers directly from application code. The server might be an open-source package pulled from a public registry, or an internal service someone stood up for a prototype that quietly became permanent.
Cloud environments
Teams deploy MCP servers into cloud accounts to support agents in staging or production. If those accounts sit outside the main security review process, so do the servers.
The MCP security risks shadow servers create
An approved MCP server can still be risky. An unapproved one carries every MCP risk with none of the controls. The OWASP MCP Security Cheat Sheet and vendor guidance point to several that shadow servers make worse.
Broad, long-lived credentials
Shadow servers often run on a personal access token or API key with far more scope than the task needs. OWASP lists excessive permissions and over-scoped tokens as a core MCP risk, because one compromised token can reach everything that user can.
Tool poisoning and rug pulls
Models read tool descriptions to decide what to call. OWASP describes tool poisoning, where hidden instructions in a tool’s description steer the model, and rug pulls, where a server changes its tool definitions after someone trusted it. An unreviewed server never had its descriptions checked in the first place.
Supply chain exposure
Many shadow servers are installed straight from public package registries or GitHub. OWASP flags installing unreviewed or compromised server packages as a supply chain risk, including typosquatted package names.
Untrusted third-party servers
Anthropic’s own guidance on custom connectors warns that they can reach “arbitrary services that have not been verified,” and that malicious servers may hide instructions that push the model toward unintended actions.
No audit trail
When an MCP server runs on a laptop or inside one team’s code, there’s no central record of which agent called which tool, with what data, and what happened. If something goes wrong, the investigation starts from nothing.
How to find shadow MCP servers
No single signal finds them all, because shadow MCP hides in different layers depending on how it was adopted. A practical discovery program looks in several places at once.
| Where to look | What you’ll find |
|---|---|
| Endpoints | Locally installed AI tools and the MCP servers they start |
| Code repositories | MCP server packages, client code, and agent configurations before they reach production |
| Identity provider logs | Sign-in and authorization patterns linked to AI tools and the users behind them |
| Network and API traffic | Outbound calls from agents and AI tools to MCP servers and model providers |
| Cloud accounts | MCP servers and agents deployed in cloud environments |
Two principles make the results useful. First, attach context to everything you find: who owns it, what it can reach, and what data it touches. A list of server names without owners is hard to act on. Second, keep the inventory current. Developers add servers every week, so a one-time audit is out of date almost as soon as it’s done.
How to bring shadow MCP servers under control
Banning MCP outright rarely works. People adopted these servers because they made work faster, and they’ll find a workaround. The goal is to make the approved path easier than the shadow one.
- Inventory first. Build a live record of every MCP server in use, with its owner, credentials, and the systems it touches.
- Triage by risk. Prioritize servers with write access, production data, or broad tokens. A read-only documentation server is less urgent than one that can push code.
- Offer an approved equivalent. For each high-value shadow server, provide a reviewed version through a central gateway, so users keep the capability.
- Scope access by role and tool. Give each team only the tools, and where possible the parameters, it needs.
- Scan tool definitions. Check tool names and descriptions for hidden prompt injection before they reach a model.
- Log every call. Route MCP traffic through a gateway that records every tool call, including denied and failed ones.
- Retire what’s left. Once an approved path exists, remove shadow servers and revoke the credentials they used.
How Airia helps
Airia covers both halves of the problem: finding shadow MCP servers and giving teams a governed alternative.
Airia Discover builds an AI inventory using nine discovery signals across workforce, infrastructure, and application and code environments. MCP server detection and code repository scanning are among them, alongside endpoint, identity, browser, network, and API signals. It doesn’t require a browser extension or SASE integration, though it supports both. A first inventory comes back within 24 to 48 hours and stays current from there. Airia consistently finds two to four times more AI than organizations expect.
Every MCP server Discover finds carries its owner, access scope, and the data it touches, and risk classifications assign each one a risk tier so teams know where to start.
From there, the Airia MCP Gateway provides the approved path. Admins approve which servers can be used, assign access at the app, tool, and parameter level by role, and serve every user through one gateway address. Every tool’s name and description is scanned for hidden prompt injection before it’s exposed, and every tool call is logged, whether it succeeded, was denied, or errored. AI guardrails and agent constraints apply the same policies to the agents using those tools.
Conclusion
Shadow MCP servers aren’t a sign that people are careless. They’re a sign that MCP is useful and that the approved path hasn’t kept up. Find them with more than one signal, triage by what they can reach, give teams a governed alternative, and keep looking. The servers nobody approved are the ones most likely to surprise you.
Find the MCP servers nobody approved. Request a demo to get a first AI inventory in 24 to 48 hours and see how Airia moves teams onto a governed gateway.
Put these ideas to work.
Schedule a 30-minute walkthrough with our team.