MCPLast reviewed

MCP server security: the risk and how to rank controls for it

MCP servers give coding agents tools: databases, ticket systems, cloud consoles, browsers. Each one widens what an agent can do with the developer's access. Here is what makes them risky, how each approach handles MCP connections, and a checklist for reviewing any server.

Why MCP servers are a risk

It runs with the developer's permissions

A local MCP server is a program on the developer's machine. Whatever the developer can read, write or reach, it can too.

Anyone can publish one

Servers come from registries, package managers and repositories. Unofficial copies of popular servers exist, and a config line can fetch and start the latest version each time.

Its tool descriptions talk to the model

A server describes its tools in text the model reads. A malicious or careless description can steer the agent, and a tool's response can carry injected instructions.

It changes without a review

A server approved last month may ship new tools or new behaviour today. Unpinned versions update on the next start.

Nobody has the list

Servers are configured per developer, per agent, per project. Without discovery, the security team cannot say which are connected.

How each approach handles MCP connections

For MCP, four of the six criteria decide the outcome: visibility, coverage of terminal and MCP actions, enforcement and audit evidence. This table orders the approaches by those four scores alone (out of 12).

ApproachDiscovers serversControls callsRecordsMCP score
Agent access security brokers (AASB)
Listed: Unbound
Finds every MCP server connected to every agent, official or not, and who installed it.Allows, blocks or requires approval per team, per server or per tool call.Logs each MCP call and the decision on it.12/12
Built-in agent settingsLists the MCP servers configured in that one agent.Allow or deny servers and tools by name in that agent's settings.Whatever the agent keeps locally; not central.6/12
IAM and secrets controlsDoes not see MCP servers.Limits what credentials given to an MCP server can reach.Provider audit logs of credential use, attributed to the identity.6/12
AI gatewaysSees MCP tools only as tool definitions inside model requests.Can strip or block tool calls in model responses, without the server's context.Model request logs; not the MCP call or its result.6/12
Endpoint detection and response (EDR)Sees local MCP server processes as processes, not as MCP servers.Can kill or block a process it judges malicious.Process and network telemetry.4/12
AppSec scanningCan find MCP configuration files committed to repositories.None at runtime.Scan results for code and configuration.4/12
CASB and DLPMay see remote MCP endpoints as web destinations; does not see local servers.Can block a remote MCP domain outright.Web and data event logs.4/12

Ties keep the order of the main ranking.

Checklist for reviewing an MCP server

Run it before a server is allowed, and again when its version changes. A broker can enforce the answers across the fleet; without one, they live in a document and each developer's config.

Enforcing it across every agent

Unbound discovers every MCP server on every agent, flags unofficial ones, and sets policy per team, per server or per tool call, with each call recorded.

Find your MCP servers with Unbound
  1. Who publishes it?The official server from the vendor of the system it connects to, a known maintainer, or an unknown copy? Check the source repository and package owner.
  2. Is the version pinned?A fixed version or digest, not a command that fetches the latest on every start.
  3. What can its tools do?List the tools. Mark any that write, delete, execute commands, send messages or move money.
  4. What credentials does it get?A scoped token for this purpose, or the developer's full session? Prefer read-only and short-lived.
  5. Local or remote?A local (stdio) server runs on the machine; a remote one sends data to someone else's endpoint. Know which, and where that endpoint is.
  6. What data leaves through it?Can tool calls carry source code, secrets or customer data out of the company?
  7. Which agents and teams may use it?Allow it for the teams that need it, not everyone by default.
  8. Which calls need a person's approval?Destructive and outbound tools should wait for a yes, not run on auto-approve.
  9. Is every call recorded?Server, tool, arguments in summary, agent, user and the policy decision, kept centrally.
  10. When is it reviewed again?On every version change, and on a schedule. Remove servers nobody has used.