verifiedagents.ai
All posts

5 min readAI Agents · Cybersecurity

How do you test whether your MCP servers are under control?

By Josh Woodruff, Founder & CEO, MassiveScale.AI | Founding Chair, Agentic Trust Framework at the CSAI Foundation

Hero: How do you test whether your MCP servers are under control?

TL;DR: Test MCP control in two parts. First, count your MCP servers twice, once from the platform team and once from developers. The counts should match. Second, score every server on five checks, from a registry row to logs the agent can't edit. Any server that can write to production and scores under five fails.

Sixteen security teams raised the Model Context Protocol on my calls between February 16 and August 28, 2026. The early ones wanted a definition. The later ones already had servers running and wanted to know one thing. Are these under control? That question has a test.

What makes MCP hard to keep under control?

MCP is a standard way for an AI model to reach tools and data. A server publishes a set of actions, and any agent that speaks MCP can call them.

The hard part is who decides. With a normal API, a person or a program decides to make the call. With MCP, the model picks the action while it's running. So the list of things an agent can do stops matching the list of things it will do.

Servers also arrive by accident. A developer installs a coding assistant that ships with MCP turned on. Now that laptop has a path into internal systems, and nobody filed a ticket. The fuller story is on MassiveScale.AI, in why security teams now ask about MCP before they roll it out.

How does the assessment score this?

Question 20 of the free assessment asks how agent actions are constrained. It's part of Segmentation, one of the five ATF elements. For MCP, each answer looks like this.

Answer

What it says

What it looks like on an MCP server

A

No action boundaries. Agents can do anything.

Every agent can call every action the server publishes.

B

Informal limits based on team guidelines.

A wiki page says which actions to avoid. Nothing enforces it.

C

Documented action allowlists.

Each agent has a written list of allowed actions.

D

Enforced action limits, with exceptions logged.

A call off the list fails and leaves a log entry.

E

Policy-as-code with real-time evaluation, which means code checks each call as it happens.

Each call is checked against rules before it runs.

How do you count MCP servers?

Count them twice.

Ask the platform team how many MCP servers are running. Then ask the developers the same thing, separately. The two numbers probably won't match. The difference between them is your first finding.

Then check what came bundled. Coding assistants often ship with MCP support. Anyone who installed one may have opened a path without meaning to.

What are the five checks for each server?

Score one point for each yes.

  1. Is it in a registry with an owner's name? One row per server, with what it reaches.

  2. Does it have its own credential? A token shared across servers means one break-in reaches all of them.

  3. Does each agent have a tool allowlist? The agent gets the actions its job needs, and the rest of the catalog stays closed.

  4. Do write actions wait for approval? Reads can run free. Anything that changes a record waits for a person.

  5. Are the logs out of the agent's reach? If the agent can edit its own log, your evidence lives inside the thing you're checking.

The pass bar is strict where it counts. Any server that can write to production needs five out of five. A read-only server on a wiki can pass with the first two.

Check 2 has outside backing. The MCP security best practices say a server must not accept a token that wasn't issued for it.

What does a failed check look like in real life?

A security leader at a healthcare payments network walked me through his worry. His shop could reach hundreds of tools through one cloud provider's MCP integration. He wasn't afraid of an attacker. He was afraid of an accident: an agent stuck in a loop that deletes everything.

He asked for "a roadblock in the way, to prevent accidental misuse." That's check 4 in his own words. He didn't want an alert after the fact. He wanted a stop. I cover how to prove that stop works in testing whether you can stop an AI agent before it acts.

Frequently asked questions

Is MCP a security risk by itself?

No. It's a standard for connecting models to tools. The risk comes from what you connect and how tightly you scope it.

Does an API gateway cover MCP?

Partly. A gateway can check the caller and cap the rate. It can't judge whether a run of allowed calls adds up to something nobody approved. That's why checks 3 and 4 exist.

Who should own the MCP registry?

Pick the owner on purpose. On my calls, ownership kept settling with whichever team showed up first. That's a default, and it's rarely the best choice.

How often should I rerun the test?

Every quarter, and whenever a new coding assistant rolls out. To see what any one agent can reach through its servers, use the ten-minute test of what an AI agent can reach.

Key takeaways

  • Count MCP servers twice, from two sources. A mismatch is a finding.

  • Score each server on five checks, one point each.

  • A server that can write to production needs five out of five.

  • A shared token fails check 2, however well the rest is set up.

Question 20 is one of 30 in the free assessment. It takes about ten minutes and scores you on all five ATF elements.

In February the question was what MCP even is. By August the servers were live. Score the ones you have before the next wave arrives.

See where your agents stand.

The free assessment takes ten minutes and scores you on the five elements of the Agentic Trust Framework.