Mittwald, a German host that serves mostly web agencies, offered its customers an MCP server: the piece that lets an AI assistant work inside their hosting account, listing projects, changing settings, doing what the customer could do. On 4 November 2026 Mittwald found that someone had been using a flaw in that server since 27 October to read customers' session tokens, and with them, to walk into 56 customer projects. A week later the attackers were back, this time through Mittwald's own website.

It is the first breach of a hosting provider through its AI-assistant interface that this site is aware of. The lesson is older than AI: a service that acts on behalf of many users holds all of their keys in one place.

[INFO]
Primary sources: Mittwald's FAQ "MCP-Server: Was du jetzt wissen musst" (last updated 7 November 2026) and its status page, incident 255116 (5 to 13 November), both read in German on 14 November. The number of projects and Mittwald's remarks on AI and the authorities are from The Record (9 November). Mittwald has promised a postmortem; it was not out when this was written.
◈ interactive artifact
The Server Between You and Your Assistant
An AI assistant working in a hosting account, the MCP server it talks through, the tokens that server holds, Mittwald's own status updates, and Mittwald's checklist.

//What Happened

Mittwald's FAQ gives the outline. The MCP server was the interface for connecting mStudio, Mittwald's hosting platform, to AI assistants. Through a vulnerability in it, the attacker read session tokens. Mittwald says no passwords were read and all affected tokens have been revoked. A token gave access to mStudio "with the user's rights", including every project that user could reach. Unauthorised access began on 27 October; Mittwald found it through its own analysis on 4 November. The attacker's preparation was disguised as normal use, which is why it went unnoticed for about a week. The Record, quoting Mittwald, puts the toll at 56 customer projects and says the attacker first tested the flaw earlier in October.

The status page shows the response. On 5 November at 11:34 Mittwald disabled mStudio logins, cut and blocked SSH, and switched off extensions. That afternoon it deactivated "a small number of projects" and emailed their owners, and at 17:02 it confirmed unauthorised access to mStudio via the MCP server. mStudio and SSH came back on 7 November, with passkeys newly available. The MCP server stayed off, and Mittwald says whether it returns will be decided after the investigation. It points customers to its command-line tool instead. Mittwald told The Record it had informed the relevant authorities.

//The Second Wave

At 02:51 on 12 November Mittwald restricted SSH again and took its website offline. At 09:18 it said the attackers had "resumed their attacks on our infrastructure" and had exploited a vulnerability to compromise the company website. It had found no access to internal systems, but shut down the customer centre, mStudio and all SSH connections as a precaution. Hosted websites and mail stayed up. On 13 November mStudio, SSH and the customer centre were restored by midday, with a captcha and email-based two-factor authentication now on logins, and the website was back at 14:04. Mittwald says there is no remaining unauthorised access.

Mittwald has not said the second wave used the MCP flaw, and the MCP server was already off, so they should be read as two incidents. Mittwald's own wording, that the attackers had "resumed", suggests it believes the same people are behind both.

//Why the MCP Server Was Worth Attacking

An MCP (Model Context Protocol) server turns a service's functions into tools an AI assistant can call. To call them as you, it needs your credentials or a session that stands in for them. Run one for a whole customer base and it holds a live session for everyone who has connected an assistant. A flaw that lets an outsider read that store is worth as much as a flaw in the login system, and it sits in a newer, less tested component.

That is the same shape as the other provider breaches this autumn. DreamHost's web File Manager received every customer's FTP password because it had to log in for them. cPGuard's update channel ran as root on every customer server because it had to update them. Mittwald's MCP server held every connected customer's session because it had to act for them. Each was a convenience layer that had to be trusted with what it served, and each became the way in. The AI part is new; the pattern is not.

//Was It an AI Attacker?

Mittwald thinks so. Its spokesperson told The Record that the logs show the attacker "worked his way through our system extremely quickly", with "only seconds" in places between steps that would take a person much longer. The Record points out that Mittwald offered no other evidence and that ordinary scripts are fast too; Mittwald did not say how it told them apart. Treat it as Mittwald's assessment, not a finding. It would not change the advice either way: after the WorkOS breach, this is the second incident in a month where the speed of an automated attacker mattered less than the gap it walked through.

//What Customers Should Do

Mittwald's FAQ says anyone who used the MCP server on or after 27 October should assume compromise and: change their mStudio password; turn on two-factor authentication; check API keys, project tokens and SSH keys, delete unknown ones, and look for SSH users they did not create; renew third-party credentials stored in projects (databases, email, external APIs); and tell the other users on the project. Customers who never used the MCP server are not affected, Mittwald says, though it recommends two-factor authentication for everyone.

The fourth step matters most. The token opened the project, and the project holds other secrets. Changing the hosting password does nothing for the database password in the site's config file or the payment-provider key in its environment.

//If You Run an MCP Server

Give the server the narrowest credential that does the job: scoped, short-lived tokens per tool rather than a full user session, and per-user isolation so one request cannot reach another user's tokens. Treat the server as part of your authentication system, with the same review and testing. Log every tool call with the user it acted for, and alert on volume and speed, since an attacker working through tokens looks like many users doing many things very quickly. And have the off switch Mittwald used: the ability to turn the whole interface off without taking the platform down with it.

[IOC]
Organisation: Mittwald CM Service (German web host); component: MCP server for the mStudio platform
Window: flaw tested in October; access from 27 October 2026; found 4 November; services restricted from 5 November; second wave 12 November (company website compromised); restored 13 November
Accessed: session tokens (no passwords, per Mittwald); 56 customer projects (The Record)
Status: MCP server off indefinitely; vulnerability not published; postmortem promised

//Sourcing and Uncertainty

The MCP server's role, the token theft, "no passwords", the rights a token carried, the dates, the response and the customer steps are from Mittwald's FAQ and status page, translated from German by this site. The 56 projects, the October testing, the AI quotes and the notification of authorities are from The Record. Exact times for the attacker's first test and first access circulate in secondary coverage; this site could not find them in a primary source and does not use them.

What is unknown. The vulnerability, and whether it was in Mittwald's code or an MCP library; what the attacker did inside the 56 projects; how the 12 November website compromise happened; and whether the two waves are the same actor.