Shadow MCP servers are connections used without the organization’s review, approval or inventory. A developer may add one through a local configuration file or a project repository. If that route bypasses your normal review process, you can end up with access to real systems that the responsible team has not recorded.
If you have read what MCP is and why its security matters, you know the risk of a single server. This is the organizational version, and it is harder, because you cannot evaluate what you cannot see.
Reviewing my own AI connections showed me how incomplete a mental inventory can be. Writing this post, I opened the connector settings in my own AI client and found three active connections: Google Drive, Notion, and a browser extension that can drive my browser directly. I had approved every one myself, one at a time, without thinking of them as a set. A connector label alone does not establish that it uses MCP; that needs checking against its implementation or documentation. Asked beforehand to list what my assistant could reach, I would have guessed wrong.

Three connections, approved one at a time, reaching my documents, my notes, and my browser. Nothing here is unauthorized. I just could not have named it from memory.
This is an inventory lesson, not evidence that those approved connectors were unauthorized or malicious. For a team, the next step is to identify which connections use MCP and whether each has an owner and approval record.
Why procurement can miss shadow MCP servers
Procurement records can help find paid tools, but a free local package may leave no invoice. A local MCP server can be launched from a command in a client or project configuration. If your process does not review those configurations, purchasing controls alone will miss that route.
A stdio connection between a client and a local process does not need a new remote SaaS endpoint. It can still leave evidence in process telemetry, package records and configuration files. Outbound API calls may also appear in network or service logs. Remote MCP services can leave OAuth grants and other account records. No single source should be assumed to give a complete inventory.
How a shadow MCP server arrives through a config file
The sharpest version of this is not a developer choosing badly. It is a server no one chose at all.
In February 2026, Check Point Research disclosed a set of vulnerabilities in Claude Code that make the point precisely. Claude Code reads project-level configuration from files that live in the repository: .claude/settings.json and .mcp.json. Changes can arrive in a cloned repository, a checked-out branch or a local edit; a merged commit is not required. Branch protection and review help only where those controls cover the path being used. Researchers found that a repository could set enableAllProjectMcpServers, which auto-approves every MCP server defined in the project, and the attacker’s server would then execute before the user could even read the trust dialog. In their demonstration the payload launched on top of the still-pending prompt, and they escalated the same technique to a reverse shell.
Anthropic patched the startup trust-dialog bypass in version 1.0.111, before Check Point published its February 2026 account. The broader review lesson remains relevant: repository configuration can request executable behavior. Check Point lists three realistic paths: a malicious pull request that carries the config alongside genuine code changes, a useful-looking honeypot repository, and an internal repo poisoned by one compromised account.
The lesson is to review configuration that can start processes, not just application code. Assign that responsibility explicitly.
Version pinning helps control local package updates, but it does not freeze a remote service or every dynamically supplied description. Review package changes, tool metadata and permission changes as separate parts of the connection.
Invisible connections, measurable costs
IBM’s breach research provides context for unmanaged data and AI. It does not measure the cost of shadow MCP specifically.
In its 2024 report, 35 percent of studied breaches involved shadow data. IBM reported an average cost of 5.27 million dollars for that group and an average identification-and-containment time of 291 days. These are findings about the studied breaches, not predictions for a particular MCP deployment.
In the separate 2025 study, 20 percent of organizations reported a breach involving shadow AI. Organizations with high shadow-AI use had average breach costs 670,000 dollars higher than organizations with low or no shadow-AI use. That is a comparison between groups, not an amount to add to the global average. IBM also reported that 63 percent lacked an AI governance policy or were still developing one.
Shadow MCP can overlap with shadow AI or unmanaged data, but it does not necessarily involve both. The reports support taking an inventory seriously; they do not establish a specific MCP cost premium or prove that delay caused every additional cost.
Find which MCP servers are already connected
To make the policy concrete, first check what is running now. Start with a bounded inventory of one team or client.
MCP configuration lives in predictable places. Client config files on developer machines (Claude Desktop, Claude Code, Cursor, VS Code and similar all keep their own), and repository-level files like .mcp.json and .claude/settings.json that ship with projects. Grepping your own repositories for those filenames is a genuinely useful afternoon, and it finds the category the Check Point research describes: servers that arrive with a codebase rather than a decision.
The second question is not “how many” but what each one can reach. A server holding a read-only token against one dataset and a server holding an admin key are different entries on the same list. Record the credential owner, identifier and scope, not the secret itself.
The third is who to ask when the answer is unclear. Ask users to confirm what they installed, then compare their answers with configuration, endpoint and OAuth records. Make reporting straightforward; memory alone can miss old or inherited connections.
Set review rules for the MCP server list
Start with an owner and a record of approved servers, versions, permissions and revocation steps. Review a candidate’s publisher and dependencies, then test its behavior with non-sensitive data. A clean test run is useful evidence, but cannot prove the absence of malicious behavior.
Where you need centralized enforcement, consider an organization-controlled gateway for allowed tools, access checks and logs. It helps only if connections are required to pass through it and bypass paths are controlled.
A client may offer only a server-level toggle. That can expose more tools to the model than the task needs, although backend permissions and credentials still constrain execution. Prefer per-tool controls where available, and check what the server can actually do with the granted access.
FAQ
QWhat is shadow MCP?
What is shadow MCP?
QHow do I find out which MCP servers my team has installed?
How do I find out which MCP servers my team has installed?
.mcp.json or .claude/settings.json. Searching your codebases for those filenames finds servers that arrived with a project, then compare findings with user reports, endpoint records and remote-service OAuth grants.QCan an MCP server be added without a developer approving it?
Can an MCP server be added without a developer approving it?
QShould we just ban MCP servers at work?
Should we just ban MCP servers at work?
QIs shadow MCP different from shadow AI?
Is shadow MCP different from shadow AI?
Sources
- Check Point Research: RCE and API token exfiltration through Claude Code project files (CVE-2025-59536, CVE-2026-21852)
- IBM: 2025 Cost of a Data Breach report findings
- IBM: Hidden risk of shadow data and shadow AI leads to higher breach costs
- GitHub Security Advisory GHSA-4fgq-fpq9-mr3g (CVE-2025-59536)
Vulnerability details refer to Check Point’s February 2026 publication and the vendor’s earlier patch. IBM figures are labeled by report edition. The inventory and review steps are recommendations, not measured outcomes for MCP deployments.