German web host Mittwald CM Service admitted over the weekend that attackers compromised 56 customer projects — and went undetected on its systems for eight days. The way in was not a web server or a database but the company's Model Context Protocol (MCP) server, the component that lets customers connect AI assistants to mStudio, its platform for managing websites and hosting.
Managing director Robert Meyer said the attackers exploited a flaw in that MCP server. Whether the bug sat in Mittwald's own code or in open-source MCP libraries is still unclear. The attacker disguised his access attempts and used diversionary tactics, Meyer wrote.
The reason Mittwald believes AI was involved is worth reading closely: the logs show the intruder "worked his way through our system extremely quickly", with "in some cases only seconds between steps that would take a human attacker significantly longer". The company offered no other evidence. The Record noted that conventional scripts also move faster than an unaided human, so the AI attribution remains the operator's inference rather than a demonstrated fact.
The timeline is pointed. The attacker first tested the MCP flaw in October, in the middle of the global scramble to patch PHP2Shell. The intrusion was discovered on November 4. Mittwald disabled the MCP server, locked accounts, brought in external forensics and informed authorities; the customer centre and managed websites were not affected, the company says.
The case is no longer an outlier but a pattern. MCP servers are where an AI assistant is granted real permissions — which makes them the place where a single flaw reaches exactly as far as the assistant does. A hosting provider is a particularly bad place for that, because its customers' projects are the thing on the other side of the wire.




