Live session tokens for a hosting control panel should never sit where an integration layer can spill them into an attacker’s hands. In late October 2026, a vulnerability in Mittwald’s Model Context Protocol (MCP) server did exactly that: an attacker read session tokens from the interface that connects customer AI assistants to mStudio, then opened mStudio and customer projects with those users’ permissions. According to Mittwald’s MCP FAQ, no passwords were taken, compromised tokens were revoked after detection, and 56 customer projects went through forensic review once the connector was treated as breached.

That is the whole initial-access story in plain terms. Nobody had to guess a password or approve a prompt. The tokens were already valid proof of a signed-in mStudio user, and the MCP server was the place they were exposed.

What the MCP connector actually handed over

Mittwald runs mStudio as the management plane for customer hosting projects. MCP was the bridge so AI assistants could talk to that plane. According to the company’s own FAQ wording, “Dabei wurden Zugangsdaten (Tokens) von Nutzern dieser Schnittstelle ausgelesen.” Those were session tokens (Sitzungs-Tokens), not password hashes and not a dumped credential store of login secrets.

The same FAQ is blunt on the password question: no passwords were read out, only session tokens, and every one of those tokens was revoked so it was no longer valid. For anyone who had used the MCP path, Mittwald stated the practical effect in one line: the token allowed access to mStudio with that user’s rights. Project reach followed whatever those rights already covered.

That distinction matters for anyone still framing every hosting incident as “stolen passwords.” Password reset theater does not describe this path. A bearer session already accepted by mStudio is enough to act as the user until something kills the token or the session backend stops honoring it. Mittwald also blocked API keys as a precaution and told affected customers to hunt unknown API keys, project tokens, and SSH keys or users that might have been planted while a stolen session was live.

Kundencenter-only sites were stated as outside this incident’s blast radius. The problem sat in the MCP server path into mStudio, not in every Mittwald login surface at once.

Held by attacker Source on Mittwald’s side What it unlocked
mStudio session tokens Vulnerable MCP server mStudio as the real user
User-level permissions Token already authorized Projects that user could open
No passwords Confirmed in MCP FAQ Login secret store not the prize

The October access window before shutdown

Intake timeline notes put first attacker testing around 21 October 2026. Unauthorized access through the MCP server is treated as underway from 27 October 2026, and Mittwald marks the server compromised from that date. Detection came from Mittwald’s own analysis on 4 November 2026. The MCP server was shut down the same day. Compromised session tokens were revoked and API keys were blocked in the same containment push.

Public incident updates on 5 November 2026 describe the broader freeze that followed: mStudio login, SSH, extensions, and related functions disabled as a precaution, some projects temporarily deactivated, forensic review of affected projects underway. Hardening continued on 6 November 2026 with system reboots, further restrictions such as orders, and work to prepare mStudio re-enablement. By 7 November 2026, mStudio and SSH access were restored, some precautionary website blocks remained, and passkey setup was offered as mStudio came back.

Eight calendar days from the 27 October start of unauthorized MCP access to the 4 November shutdown is a long time to leave a token-reading integration online. Mittwald’s FAQ places the hole clearly: “Die Lücke lag in unserem MCP-Server.” The exact bug class is not spelled out in the public material, and no threat actor is named. What is named is enough for defenders: a server-side read of live session material on an AI-assistant connector, followed by use of those sessions against mStudio.

Fifty-six customer projects were reviewed as in scope of compromised-token access. Mittwald has not published a confirmed, estate-wide inventory of every file, database, or mailbox touched inside each project. Customers were pointed at case-by-case notification where needed, plus the standard hygiene list: change the mStudio password, turn on two-factor authentication, renew third-party credentials stored in projects, and delete unknown keys.

Why login MFA was not the control that failed here

Session tokens stolen from an application server are not a story about someone typing a one-time code into a fake page. Authentication for those users had already succeeded earlier. The MCP flaw let someone else present the leftover proof. Mittwald’s public record does not name which MFA method, if any, sat on mStudio logins at the time, and it does not need to for this mechanism. A second factor at the password prompt does not stop a vulnerable connector from reading a token that the panel already issued.

Official guidance still told customers to enable two-factor authentication and change passwords. That is sensible hygiene after any control-plane scare, and it shrinks reuse if passwords or recovery paths were ever weak. It is not a description of how this attacker got in. The in-path failure was token exposure on the MCP server, then replay of those sessions into mStudio until revocation.

Containment matched that reality. Mittwald invalidated the stolen sessions, took MCP offline indefinitely pending investigation, briefly locked large parts of mStudio and SSH, and only then restored service with tighter posture and passkey setup available. Passkeys on a restored login path improve future sign-ins. They do not rewrite the October window in which bearer tokens were already readable from the connector.

Hosting teams watching this from the outside should separate three jobs that often get mashed together in incident mail. Job one is stopping the server bug that can emit session material to anyone who finds it. Job two is making stolen sessions die fast when misuse shows up: revoke, rotate API and project tokens, kill odd SSH users. Job three is hardening how new logins are proven going forward. Confusing job three with job one is how postmortems end up arguing about OTP apps while the real hole was an integration holding live Sitzungs-Tokens.

The same pattern will keep showing up wherever AI connectors, admin APIs, and long-lived browser or API sessions share a trust boundary. If the connector can see a token that mStudio will honor, the connector’s vulnerability budget is effectively the control plane’s vulnerability budget. Mittwald said the quiet part out loud by pulling MCP down hard rather than leaving an assistant bridge online “for convenience” during forensics.

A related mfa2point0.com article covers session token theft after a legitimate mStudio login. session token theft after a legitimate mStudio login.

FAQ

Were passwords stolen in the Mittwald MCP incident?

No. According to Mittwald’s MCP FAQ, passwords were not read out in the Mittwald MCP session token theft. Only session tokens (Sitzungs-Tokens) from users of the MCP AI-assistant interface were taken, and Mittwald states those tokens were all revoked and are no longer valid.

How did attackers reach mStudio customer projects?

Attackers reached mStudio customer projects by using stolen session tokens from Mittwald’s vulnerable MCP server as if they were the legitimate user. Mittwald’s FAQ states each such token allowed access to mStudio with that user’s rights, which extended to projects those users could already open. Fifty-six customer projects were forensically reviewed as in scope of that access.

When did unauthorized MCP access start and when was it stopped?

Unauthorized access via Mittwald’s MCP server is treated as beginning on 27 October 2026, with earlier testing noted around 21 October 2026. Mittwald detected the activity through its own analysis on 4 November 2026 and shut down the MCP server the same day while revoking compromised session tokens and blocking API keys as a precaution.

Did multi-factor authentication stop or fail in this breach?

Multi-factor authentication at login was not the gate the attacker had to pass in the Mittwald MCP incident as disclosed. Session tokens were read from the MCP server after users already had valid mStudio sessions, so the attacker replayed existing proof of access rather than completing a fresh password-and-second-factor challenge. Mittwald has not named the MFA type in use at the time in the public FAQ.

What did Mittwald tell affected MCP users to do afterward?

Mittwald told affected MCP users that compromised session tokens were revoked, advised changing the mStudio password, enabling two-factor authentication, reviewing and deleting unknown API keys, project tokens, and SSH keys or users, and renewing third-party credentials stored in projects. mStudio and SSH were later restored after precautionary lockouts, with passkey setup made available at restoration.