Mittwald breach exposes 56 customer projects through a flaw in its AI connector
German hosting provider Mittwald said attackers exploited a vulnerability in its Model Context Protocol (MCP) server, which connects AI assistants to its mStudio management platform, to access 56 customer projects. The intruder went undetected for eight days; Mittwald said passwords were not accessed and its scans found no further activity. Managing director Robert Meyer suggested the attacker used AI tools based on the speed of their actions, but The Record reported that Mittwald provided no direct evidence.
Bottom line — Mittwald revoked session tokens and is investigating what project data was accessed.
Go deeper 6
-
The Record says the flaw’s location is unclear: it may have been in Mittwald’s code or in open-source MCP libraries.
-
Mittwald disabled the affected MCP server, revoked compromised session tokens and brought in external forensic experts, according to BeyondMachines.
-
BeyondMachines reports that Mittwald notified relevant authorities and restored mStudio and SSH access on November 7.
-
The Record says the attacker first tested the flaw on October 21 and exploited it on October 27, while Mittwald discovered the intrusion on November 4.
-
DEV Community notes that speed alone does not establish AI use; conventional scripts can also move through systems quickly.
-
AuthZed’s breach timeline lists recurring MCP risks including over-privileged tokens, untrusted content influencing tool calls and weak separation between connected systems.