What Is Model Context Protocol (MCP)?

Definition: Model Context Protocol (MCP) is an open standard that lets AI assistants and agents connect to external tools, data sources and services through one common interface.

Definition: Model Context Protocol (MCP) is an open standard that lets AI assistants and agents connect to external tools, data sources and services through one common interface. An MCP server exposes tools, such as search a database or create a ticket, and an MCP client inside the AI application discovers and calls them on the user’s behalf.

Why it matters

MCP has quickly become the default way to plug AI agents into business systems. Anthropic open sourced it in November 2024, and in December 2025 donated it to the Agentic AI Foundation under the Linux Foundation, so no single vendor controls it. Major AI platforms and coding assistants now support it.

That convenience is also the risk. Every MCP server is a doorway from an AI model into your systems. If the server, its tools or its credentials are weak, an attacker who can influence the model can reach whatever the server can. That makes MCP a non human identity and access problem as much as an AI problem.

How MCP works

  1. Host: the AI application the user works in, such as a chat assistant or coding tool.
  2. Client: the part of the host that connects to MCP servers and passes requests between the model and the tools.
  3. Server: a program that exposes tools, data (resources) and prompt templates for a system such as a CRM, file store or database.
  4. Authorization: for remote servers, the MCP specification uses OAuth 2.1, so the client gets a scoped access token rather than a password.

MCP security risks

OWASP’s MCP Top 10 (version 0.1) lists the main risks. Seven to know:

  1. Tool poisoning: hidden instructions placed in a tool’s description that the model reads and follows, but the user never sees.
  2. Token mismanagement: long lived tokens or secrets stored in config files, logs or code.
  3. Privilege creep: a server that started read only gains write and delete access nobody approved.
  4. Shadow MCP servers: servers that developers run on laptops or test machines without security review, similar to shadow AI.
  5. Weak authentication: servers that accept requests without verifying who or what is calling.
  6. Supply chain: installing community servers or packages that are compromised or change after approval.
  7. Context over sharing: data from one user, task or customer leaking into another’s session.

How to secure MCP

  1. Keep an inventory: record every MCP server, its owner, the tools it exposes and who can use it, in the same agent registry as your AI agents.
  2. Give each server its own identity: use OAuth 2.1 with scoped, short lived tokens instead of shared API keys, following OAuth 2.0 best practice.
  3. Apply least privilege: expose only the tools an agent needs, and prefer read only access. Review scopes regularly.
  4. Approve before you install: allow only reviewed servers, pin versions, and check tool descriptions again whenever they change.
  5. Require a human for risky actions: deleting data, sending external messages or moving money should need confirmation.
  6. Log every tool call: send MCP activity to your SIEM under each agent’s own identity.

MCP vs a traditional API integration

AspectTraditional API integrationMCP
Built forOne app calling one APIAny AI agent calling many tools
Who decides which call to makeDeveloper’s codeThe AI model, at run time
Typical credentialAPI key or service accountOAuth 2.1 token per server (for remote servers)
Main new riskLeaked keysTool poisoning and model manipulation
Review neededAt build timeAt build time and whenever tools change

Example

An engineering team connects its coding assistant to an MCP server for the company’s ticketing system. Without controls, the server uses an admin API key and can close or delete any ticket. With controls, the assistant gets an OAuth token scoped to read and comment on the developer’s own project, the token expires within hours, deletions need a human click, and every call is logged under the assistant’s own AI agent identity.

FAQ

Is MCP secure? MCP itself includes an OAuth 2.1 based authorization model, but security depends on how each server is built and deployed. Most real problems come from over scoped tokens, unreviewed servers and poisoned tool descriptions.

Who owns MCP? Anthropic created MCP and donated it in December 2025 to the Agentic AI Foundation, a fund under the Linux Foundation. It is governed as an open, vendor neutral standard.

What is an MCP server? A program that exposes tools and data from a system, such as a CRM or database, so AI agents can use them through MCP.

What is tool poisoning in MCP? An attack where malicious instructions are hidden in a tool’s description or output. The AI model reads them as trusted context and may act on them without the user knowing.

Related terms: Agentic AI Security; Non Human Identity (NHI); OAuth 2.0; Prompt Injection; Shadow AI

Call to action: Connecting AI agents to your systems through MCP? Our free AI Risk Assessment reviews the servers, tokens and access first: https://sennovate.com/ai-risk-assessment/