Mittwald, a web hosting provider based in Espelkamp, Germany, has disclosed that an attacker gained access to customer projects by exploiting a vulnerability in its MCP (Model Context Protocol) server, the interface customers used to connect AI assistants to the company’s mStudio management platform. Forensic analysis identified 56 compromised projects.
What happened
According to Mittwald’s incident write-up, the attacker first probed the MCP server on 21 October 2026, then spent several days behaving like an ordinary customer, registering test accounts and trying simple injections. On 27 October, minutes after registering another account, the attacker exploited the flaw; Mittwald now treats all affected data from that date as compromised. German coverage reports that session tokens were read via the vulnerability.
The intrusion was detected on 4 November. Mittwald blocked the attacker’s accounts, shut down the MCP server, and, on 5 November, locked projects showing attacker traces, paused new registrations and temporarily disabled mStudio logins and SSH access for all customers. External forensic specialists were engaged and authorities notified. mStudio and SSH access were restored on 7 November; websites remained reachable throughout, apart from projects blocked as a precaution.
Born’s IT blog and others reported a second, separate incident on the night of 11 to 12 November in which the company website was compromised; Mittwald said it had no indication of access to internal systems or customer projects in that event.
Who is affected
Primarily customers who connected mStudio to AI tools via the MCP server; they are being contacted individually. The platform-wide lockout, however, disrupted many customers who never used the feature.
Recommendations
- Mittwald customers should rotate SSH keys, database passwords, application secrets and API tokens for any project managed through mStudio, and set up passkeys.
- Review project file systems and cron jobs for unfamiliar changes made after 27 October.
- Providers exposing MCP endpoints should apply strict authorisation per tool call, bind tokens to sessions and scopes, and rate-limit or vet newly registered accounts.
- Treat AI integration interfaces as privileged APIs subject to penetration testing before launch.