Fifty-six hosting projects became reachable because someone else held a still-valid mStudio session, not because anyone completed a fresh sign-in. In the Mittwald MCP session token theft, a vulnerability in the Model Context Protocol server that bridges customer AI assistants to mStudio let an attacker read session tokens and open mStudio with those users’ rights. According to Mittwald’s MCP FAQ, no passwords were taken, the stolen tokens were revoked after detection, and forensic review covered 56 projects. Login MFA never faced this path. The proof of access already existed on the connector.

What Mittwald confirmed about the MCP spill

Mittwald’s public FAQ states the hole in plain German: the gap sat in the MCP server, and access data in the form of tokens for users of that interface were read out. Those were session tokens (Sitzungs-Tokens), not a password dump. The same FAQ answers the password question directly: no passwords were read out, only session tokens, and every compromised token was revoked so it was no longer valid.

For anyone who had used the MCP path, Mittwald described the effect in one line. The token allowed access to mStudio with that user’s rights. Project reach followed whatever those rights already covered. Kundencenter-only sites were stated as outside this incident. API keys were blocked as a precaution, and customers were told to review unknown API keys, project tokens, and SSH keys or users that might have been added while a stolen session was live.

Unauthorized access through the MCP server is treated as underway from 27 October 2026, with earlier testing noted around 21 October 2026. Mittwald’s own analysis detected the activity on 4 November 2026. The MCP server was shut down the same day. Public updates on 5 November 2026 describe precautionary freezes on mStudio login, SSH, extensions, and related functions, plus temporary project lockouts while forensics ran. Service restoration followed on 7 November 2026, with passkey setup offered as mStudio came back.

The step-by-step access window and connector mechanics are covered in the companion breakdown on how Mittwald’s MCP server exposed live session tokens.

Why a second factor never saw this attack

mStudio had already accepted those users earlier. The MCP flaw did not sit on a login form. It sat on a server-side bridge that could see tokens the panel would still honor. Presenting a stolen session token is replay of leftover proof, not completion of a password-plus-code ceremony.

Mittwald has not named which multi-factor method, if any, protected mStudio logins at the time. That omission does not change the mechanism. A one-time code, push approval, or SMS step only runs when someone is proving identity at sign-in. Once mStudio has issued a bearer session and an integration can read that session material, the second factor from the original login is already in the past.

Official aftercare still told customers to change the mStudio password, enable two-factor authentication, and renew third-party credentials stored in projects. That hygiene shrinks reuse and plants a stronger gate on the next interactive login. It is not a description of how the attacker obtained mStudio rights in October. The in-path failure was token exposure on the MCP server, then use of those sessions until revocation.

Stage What the attacker held Did a login factor apply?
Earlier real sign-in Legitimate user session created Yes, at that earlier moment
MCP server read Live Sitzungs-Tokens No fresh challenge
mStudio replay User-level project rights Token already authorized
Containment Revoked tokens, MCP offline Session kill, not MFA

Eight calendar days from the 27 October start of unauthorized MCP access to the 4 November shutdown is a long window for any integration that can emit control-plane sessions. The exact bug class is not detailed in the public FAQ, and no threat actor is named. Defenders still have a clear job split: fix the server that can spill session material, kill stolen sessions fast, and only then argue about how the next login should be proven.

What phishing-proof sign-in changes next time

Device-bound, origin-bound authentication removes transferable secrets from the interactive login path. There is no password to steal, no OTP to relay, and no push to fatigue when the private key never leaves the enrolled device and each signature is bound to the real origin. That is prevention, not detection: the remote kit never completes a fake login in the first place.

That property does not rewrite a server process that can already read a live session after a legitimate login. The Mittwald MCP session token theft is the residual class. The attacker did not need to finish mStudio authentication. The connector handed over tokens that authentication had already produced. Phishing-proof MFA 2.0 would not have stopped that read any more than an OTP app would have. The disclosure does not support a broader claim.

What stronger sign-in still buys after an incident like this is narrower and still real. Future interactive logins stop being phishable ceremonies. Stolen passwords and coached codes stop being enough to mint new sessions. Attacker enrollment of a new factor gets harder when onboarding itself refuses transferable secrets. None of that un-steals a token that already existed in MCP memory in late October 2026.

Containment matched the actual failure mode. Mittwald invalidated compromised sessions, took MCP offline indefinitely pending investigation, briefly locked large parts of mStudio and SSH, blocked API keys, and only then restored service. Shortening session lifetime alone would not have prevented the spill. It would only have forced faster re-use or re-theft while the vulnerable connector stayed online. Revoke and connector isolation were the controls that ended the window.

Hosting and SaaS teams should treat AI connectors, admin APIs, and long-lived browser sessions as one trust budget. If the assistant bridge can see a token mStudio will honor, the bridge’s vulnerability is effectively the control plane’s vulnerability. Pulling MCP down hard was the operational answer. Leaving an assistant bridge up “for convenience” during forensics would have kept the same blast radius open.

Control Against MCP token read Against next interactive login
Password + OTP/push No effect after issue Still phishable
Passkeys at login No effect on server spill origin-bound ceremony
Full-lifecycle phishing-proof MFA No effect on this spill No transferable factor to coach
Token revoke + connector shutdown Ends the live window N/A

Passkeys harden the restored mStudio gate

When mStudio returned, Mittwald made passkey setup available. A passkey is a device-held key pair. The private key stays in hardware. The server stores only a public key. At sign-in the device produces an origin-bound signature for that identity provider’s real login origin. On a generic enterprise IdP that issues only hardware-bound signatures, a look-alike host cannot obtain a usable assertion, because the authenticator will not sign for the wrong origin.

That property is origin-bound at authentication. It hardens the login act. It does not, by itself, erase every weak recovery path or every server-side session store. For Mittwald customers, passkeys raise the bar on the next person who tries to open mStudio through a normal login. They do not describe how the MCP tokens were obtained, and they do not replace revoke when a bearer session has already leaked from an integration.

Recovery was never the story Mittwald told

Nothing in the public FAQ describes helpdesk re-enrollment, a Temporary Access Pass, or a coached recovery secret as the entry path. The stolen material was already-issued session tokens on the MCP interface. Enrollment and recovery policy still matter for the next incident class, where an attacker talks support into minting a new factor. Full-lifecycle phishing-proof coverage closes registration, device onboarding, authorisation, authentication, and decommissioning so no phishable OTP, SMS code, or email link becomes the onboarding gate. That is a different failure mode from a connector that can read Sitzungs-Tokens after a finished login.

Mittwald’s guidance to enable two-factor authentication and change passwords after restoration is still sound hygiene. Pair it with the harder lesson from this disclosure: do not park live control-plane sessions where an AI assistant bridge can spill them, and assume any token an integration can read is a token an attacker can replay until you kill it.

Key takeaways for defenders

The industry will keep wiring assistants into hosting panels. The teams that survive that pattern will assume every connector inherits the panel’s session risk, and they will design sign-in so the only proof worth stealing is a key that never leaves the device and never appears on an integration host.

FAQ

Would multi-factor authentication have stopped the Mittwald MCP session token theft?

No. Multi-factor authentication at login would not have stopped the Mittwald MCP session token theft as disclosed. According to Mittwald’s MCP FAQ, the attacker read session tokens from the vulnerable MCP server after users already had valid mStudio sessions, then used those tokens with the users’ rights. No fresh password or second-factor challenge stood between the stolen token and mStudio access.

Did attackers steal mStudio passwords in this incident?

No. Attackers did not steal mStudio passwords in the Mittwald MCP incident. Mittwald’s FAQ states that no passwords were read out, only session tokens (Sitzungs-Tokens) from users of the MCP AI-assistant interface, and that those tokens were all revoked and are no longer valid.

What does passkey setup after mStudio restoration actually fix?

Passkey setup after mStudio restoration hardens future interactive sign-ins to Mittwald’s panel. A device-bound, origin-bound signature replaces typed secrets at login, which is origin-bound for that ceremony. Passkeys do not undo server-side session token reads that already happened on the MCP connector in October and early November 2026.

How is phishing-proof MFA different from passkeys alone for a case like Mittwald?

Phishing-proof MFA differs from passkeys alone because it removes phishable factors across registration, device onboarding, authorisation, authentication, and decommissioning, not only at the login act. Passkeys are origin-bound at authentication. Neither control prevents a vulnerable application server from reading a session token that was already issued, which is the Mittwald MCP path. Both still matter for stopping the next coached or proxied login.

What should hosting teams change after Mittwald’s MCP disclosure?

Hosting teams should isolate AI and admin connectors from live session material, revoke and rotate tokens quickly when a connector is suspect, and move interactive panel login to hardware-bound credentials so new sessions are harder to mint remotely. Mittwald’s own response paired MCP shutdown and token revoke with later passkey availability. That ordering matches a post-authentication spill better than treating the event as a failed OTP prompt.