Mittwald's incident account describes unauthorized access through a Model Context Protocol server designed to let agent clients manage hosting resources. A failure in the authorization mapping allowed activity to cross from the caller's expected project authority into other customer project scope. The vendor identified 56 affected customer projects and…
Inspect the ClaimsCase · DiggingBeagle record
Vendor reports service restoration and project count
Mittwald reported an authorization failure in its AI-facing MCP interface that gave an unauthorized party access to 56 customer projects; its speculation about the attacker's own AI usage is not independently verified.
Public disclosure: Nov 7, 2026
A real production breach involving a hosting provider's AI-facing MCP server and the 56 customer projects it identified as affected.
Assessment profile
Assessment method- Setting
- production
- Exploitation
- confirmed in the wild
- Evidence
- Grade A
- Remediation
- fixed reported
- AI role
- AGAINST AI
Basis and provenance
Desk assessment of cited firsthand report and scope limits; quantitative denominators describe only the reporter's stated sample. No independent live exploitation or universal impact is inferred.
Assessed Nov 11, 2026 using diggingbeagle.assessment/1.
Evidence basis: CLM-R236-MITTWALD-MCP-CLIENT-PROJECT-ACCESS-01 · CLM-R236-MITTWALD-MCP-CLIENT-PROJECT-ACCESS-02 · Mittwald: security incident involving customer projects and the MCP server
At a glance
MCP provides an alternate interface into established hosting privileges. If the alternate interface bypasses project ACLs, the host can become vulnerable even though its conventional web management panel enforces access correctly.
Read the implicationsThe vendor report does not independently establish full contents taken from every project, a reusable exploit affecting other providers or actual autonomous AI direction. Project count and business/customer count are different denominators. The exact MCP endpoint and authorization predicate used in the production intrusion were not supplied in a…
Limits and uncertaintyFull account
Mittwald's incident account describes unauthorized access through a Model Context Protocol server designed to let agent clients manage hosting resources. A failure in the authorization mapping allowed activity to cross from the caller's expected project authority into other customer project scope. The vendor identified 56 affected customer projects and described containment and follow-up security work. That project figure counts affected projects, not proven data theft victims or unique companies.
The report notes speed or automation consistent with possible attacker use of AI tools, but does not provide a model transcript or causal proof that an autonomous agent planned or executed the intrusion. The affected MCP integration makes the incident intrinsically relevant to AI tooling as an attack surface. No verified field evidence justifies labeling the attacker as BY AI solely from timing.
A real production breach involving a hosting provider's AI-facing MCP server and the 56 customer projects it identified as affected.
The vendor report does not independently establish full contents taken from every project, a reusable exploit affecting other providers or actual autonomous AI direction. Project count and business/customer count are different denominators.
The vendor dates initial probing to October 21, with accounts created from October 23 and targeted exploitation on October 27. It detected anomalous activity on November 4 and temporarily disabled its AI-facing MCP server, mStudio sign-in, SSH and other account operations during containment. Its November 7 account identifies 56 affected projects, not 56 corporate victims and not a verified number of stolen documents.
Mittwald infers possible use of AI tooling from the speed of the intruder's actions, sometimes separated by seconds. That does not establish a model-directed intrusion or an ARTEX connection. The confirmed AI-specific failure was the MCP integration becoming an unauthorized path into managed hosting resources, a path ordinary dashboard users might never invoke. The company says customers' hosted websites stayed available; an organization responding to this incident should review the exact projects and access keys granted through MCP, not assume all hosting assets were reached.
Mechanism and trust boundary
- 01
Call AI-facing hosting interface
An unauthorized client invokes or reaches an MCP operation backed by provider infrastructure.
Boundary: MCP client / hosting API
- 02
Cross project authorization boundary
Server requests are dispatched or scoped to projects the caller does not own.
Boundary: tool gateway / tenant project ACL
- 03
Access affected project resources
The provider identifies 56 affected hosting projects and remediates access.
Boundary: provider API / tenant resources
Timeline
- Oct 27, 2026Event type unspecified
MCP vulnerability exploited
Mittwald places targeted abuse of an attacker-created account at 15:03 local time.
- Nov 4, 2026Event type unspecified
Mittwald discovers anomaly
First attacker account blocked and MCP server taken offline during incident containment.
- Nov 7, 2026Event type unspecified
Vendor reports service restoration and project count
mStudio login and SSH restored while compromised project review continued.
- Nov 7, 2026disclosure
Public disclosure
Claims & evidence
CLM-ENR237-MITTWALD-56-TIMELINEMittwald reports exploitation of its MCP entry point on October 27, discovery on November 4, and forensic review of 56 affected projects by November 6; it does not establish model-driven attack execution.supported
Basis: reported finding
- supportsMittwald: security incident involving customer projects and the MCP serverprimary disclosure
German first-party statement: Im Schatten von PHP2Shell and dated November 4–7 recovery timeline
CLM-R236-MITTWALD-MCP-CLIENT-PROJECT-ACCESS-01Mittwald reports an MCP server security incident involving 56 customer projects.supported
Basis: reported finding
- supportsMittwald: security incident involving customer projects and the MCP serverprimary disclosure
Sicherheitsvorfall am MCP-Server: incident scope and affected projects
CLM-R236-MITTWALD-MCP-CLIENT-PROJECT-ACCESS-02The vendor's AI-tool hypothesis is not a verified model-to-tool execution trace; the confirmed AI-security role is the MCP server as the attacked interface.supported
Basis: reported finding
- supportsMittwald: security incident involving customer projects and the MCP serverprimary disclosure
Incident account: technical cause and attribution caveats
Implications
MCP provides an alternate interface into established hosting privileges. If the alternate interface bypasses project ACLs, the host can become vulnerable even though its conventional web management panel enforces access correctly.
Controls and mitigations
- Bind each MCP operation to authenticated customer identity, allowed project IDs and the specific permitted effect at execution time.
- Do not use shared provider service identity as a substitute for per-project authorization.
- Review MCP access logs, contact identified project owners and rotate keys only where the concrete exposure path supports it.
Unknowns and contradictions
- The vendor report does not independently establish full contents taken from every project, a reusable exploit affecting other providers or actual autonomous AI direction. Project count and business/customer count are different denominators.
- The exact MCP endpoint and authorization predicate used in the production intrusion were not supplied in a reproducible public proof; do not assert the same technique affects unrelated MCP servers.
Sources and citation
Material revision history
- Nov 11, 2026 · Published version · first publication · revision 238