AI Security
Ten questions to ask before you approve a vendor's MCP connector
German host Mittwald says an attacker came in through its MCP server, the connector customers use to link AI assistants to their accounts, and reached 56 customer projects before anyone noticed. Health systems are being asked to approve connectors like this now.
November 18, 2026 · 5 min read

An MCP server is a door into your account that was built for an AI assistant, and it can be used by anyone who finds a way through it. German hosting provider Mittwald published an account on November 7 of an attacker who exploited a flaw in its Model Context Protocol server, and said it had forensically examined 56 compromised projects.
HealthTech vendors are shipping MCP connectors to health systems right now, so that a clinician's assistant can read a schedule, file a ticket, or pull a document. This post is the list of questions I would want a hospital to ask before approving one. It stays on the defensive side. I am not describing the Mittwald flaw, and Mittwald has not published its details either.
What Mittwald says happened
Mittwald hosts websites for web agencies and runs a management platform called mStudio. According to The Record, the MCP server let customers connect their AI assistants to mStudio. The Record also notes it is unclear whether the flaw was in Mittwald's own code or in open-source MCP libraries.
Mittwald's post, written by founder and managing director Robert Meyer, gives a timeline. On the evening of October 21, while, in Meyer's words, the whole IT world was busy with the PHP2Shell vulnerability, the attacker first tested a weakness in the MCP server. From October 23 the attacker used the platform as a regular user, registering test accounts. On October 27 at 14:58 the attacker registered another account and at 15:03 exploited the MCP flaw. Mittwald says it now treats all data from October 27 onward as compromised.
Mittwald detected unusual behavior on the morning of November 4, locked the first attacker account at 11:11, identified the MCP server as the entry point, and shut it down. Over the next days it blocked affected projects, temporarily disabled new registrations, locked mStudio logins and SSH access to customer projects, and brought in outside forensic help. The Record summarizes this as 56 customer projects compromised after the attacker went undetected for eight days, which matches the October 27 to November 4 dates in Mittwald's post. When mStudio came back on November 7, Mittwald added passkey support.
Meyer also wrote that the logs showed the attacker moving extremely fast, with only seconds between some steps, which he read as a sign of AI tools. The Record points out that Mittwald gave no other evidence of AI involvement and did not distinguish it from ordinary scripting. I would treat the AI angle as Mittwald's assessment, not a finding.
Why the connector is the part to look at
A connector for an AI assistant needs credentials that can act on the user's behalf. It is usually new code, written quickly, with a broad scope because broad scope makes demos work. It sits on the internet. And because it is meant for software rather than people, the activity going through it can look like normal automation.
That combination is the reason to review it separately from the rest of the vendor's product. A vendor may have a mature security program for its web app and a connector that went from prototype to production in a quarter.
Ten questions for the vendor
Ask these in writing, and ask for evidence where it makes sense.
- What can the connector do? List every action it exposes. Read-only, write, delete, and administrative actions should be listed separately.
- What scope does its token carry? Is it limited to the actions above, or does it inherit everything the signed-in user can do?
- Whose identity does it act as? A dedicated service identity, or the user's own session?
- Where are tokens stored, and who at the vendor can read them? Are they encrypted at rest, and are they ever written to logs?
- How long do tokens live, and how are they revoked? Can our administrators revoke all connector access for our organization in one step, without calling the vendor?
- What is logged for each call? We want the identity, the action, the target record, and the time, available to us on request.
- How is the connector authenticated, and is it reachable from the public internet? If so, what limits are in front of it?
- Has the connector itself been through an independent security review or penetration test since it launched? When, and what changed as a result?
- Which open-source MCP libraries or SDKs does it depend on, and how do you track and patch vulnerabilities in them?
- If the connector is compromised, how soon will you tell us, and will you be able to tell us which of our records were touched?
A vendor with good answers will give you a short list of actions, a narrow scope, short-lived tokens, a one-click revoke, and logs you can get. A vendor that cannot answer question 1 has not thought about the connector as a separate attack surface.
What to do on the hospital side
Two things do not depend on the vendor at all.
First, keep connectors off by default. Turn one on for a defined group, for a defined purpose, with an owner who reviews it. An MCP connector enabled tenant-wide because one department asked for it is how a narrow pilot becomes a broad exposure.
Second, watch the connector's activity separately. If the vendor provides logs, ingest them. Look for volume that does not match the pilot, access at odd hours, and actions outside the expected set. Mittwald found its attacker from unusual behavior. That is the signal you want to be able to see on your own side.
Where to start
If your organization has already approved any AI assistant connectors, list them this week, with the vendor, the scope, and the owner. Send the ten questions to the vendor of the one with the broadest access. The answers will tell you whether you approved a narrow tool or an open door.