MCP Security

MCP Connectors Security Review Checklist

A proper MCP connector security review examines four layers independently: authentication and authorization, tool permission scope, data flow across trust boundaries, and runtime monitoring. Because MCP does not mandate a standard authorization flow for local servers or a standard audit log format, each of these must be verified explicitly, before deployment and again whenever the server, its tools, or the underlying MCP specification changes.

Four Review Layers

Each MCP server should be assessed across the same four dimensions, regardless of transport type or vendor.

Authentication & Authorization

OAuth 2.1 enforcement on remote servers, compensating controls on local stdio servers

Tool Permission Scope

Reviewing tool definitions and input schemas for scope creep

Data Flow & Trust Boundaries

Mapping what internal systems become reachable through granted access

Runtime Monitoring

Logging and policy enforcement to catch what static review cannot

Pre-Deployment Review Checklist

Work through this list for every MCP server before it is approved for use, and repeat it whenever the server or its tool definitions change.

  • Identify whether each MCP server uses stdio or Streamable HTTP transport, since their security models differ substantially
  • Confirm OAuth 2.1 authorization is implemented for any remote server, and document compensating controls for any local server that lacks it
  • Review every tool's name, description, and input schema for scope beyond documented business need
  • Verify the host or client presents explicit consent prompts before tool invocation, resource access, or data transmission
  • Map data flows to determine what internal systems or data become reachable through the granted tool access
  • Assess transport-layer security independently, including TLS configuration and token storage, for remote servers

Why MCP Connectors Need a Structured Review Process

Model Context Protocol connectors extend what an AI system can see and do, which means each connector introduces its own authentication model, permission surface, and data flow. Without a consistent review process, these differences get evaluated inconsistently, or not at all, before a connector reaches production.

What MCP Does and Does Not Enforce

MCP's specification defines OAuth 2.1 as the authorization flow for HTTP-based remote servers, but it explicitly leaves local stdio servers outside that requirement. It also does not standardize an audit log format across implementations. Reviewers should treat these as gaps to close with documented compensating controls rather than assumptions to leave unverified.

Transport Type Changes the Risk Profile

A stdio server running locally and a Streamable HTTP server running remotely present different attack surfaces. Confirming the transport type at the start of a review determines which controls are even applicable.

Tool Definitions as an Injection Surface

Tool names, descriptions, and input schemas are passed to the model as context. Because this metadata originates from the server rather than the host application, unvalidated or manipulated tool descriptions can be used to inject instructions, a risk generally referred to as indirect prompt injection. Reviewing tool definitions for scope beyond documented business need is therefore a security control, not a formality.

Governance Practices for Ongoing MCP Trust

A static review at deployment time is necessary but not sufficient. Sustained trust in an MCP connector depends on governance practices applied consistently over time:

  • Treat MCP server trust as non-transitive: approving one server's tools does not extend trust to other servers or downstream systems it may call
  • Require documented data flow mapping for every server as a condition of approval, not an optional artifact
  • Assign clear ownership for periodic reassessment, given that MCP's authorization mechanisms have changed materially within the past year
  • Maintain an inventory of consent and approval decisions to support audit and incident response
  • Define explicit re-review triggers, including specification upgrades, tool schema changes, and server ownership changes

Where Static Review Ends and Runtime Governance Begins

Static review confirms that a connector should be approved based on its declared authentication model, tool scope, and data flow at a point in time. It cannot detect a tool that begins behaving differently after an update, or a server whose ownership changes silently. That is the boundary where runtime monitoring and policy enforcement take over, providing ongoing verification that a connector continues to operate within its approved scope.

Common Questions on MCP Connector Review

Does MCP require authentication for all servers?

No. The OAuth 2.1 authorization flow applies to HTTP-based remote servers. Local stdio servers are explicitly exempted, leaving authentication decisions to the implementer.

Is a one-time security review sufficient for an approved MCP connector?

No. Because MCP's specification and individual server tool definitions can change, reviews should be repeated on triggers such as spec upgrades, schema changes, or ownership changes.

Can tool descriptions themselves be a security risk?

Yes. Tool descriptions and input schemas are passed to the model as context, so unvalidated server metadata can be manipulated to inject instructions, a risk known as indirect prompt injection.

Does approving one MCP server extend trust to others it interacts with?

It should not. MCP server trust is non-transitive. Each server and its downstream calls require independent review and data flow mapping.

Bring Runtime Governance to Your MCP Deployments

Static review establishes whether an MCP connector should be approved. Runtime governance determines whether it stays within scope after deployment.

Explore Runtime Governance