Skip to content
StrikeCyberStrikeCyber
Field Manual · AI platform security

Model Context Protocol (MCP) Security

Reviewed November 13, 2026AI platform security 1 min read

Quick answer

The Model Context Protocol (MCP) is an open standard for connecting AI applications and agents to tools and data sources. It makes agents far more capable, and each MCP server an agent connects to extends what an attacker could do if the agent is manipulated. MCP security is about controlling which servers are trusted, what they can do and how agents use them.

On this page
01

How MCP Works

An MCP client, such as an AI assistant or coding agent, connects to one or more MCP servers. Each server exposes tools (actions the agent can take), resources (data it can read) and prompts. The model reads tool descriptions to decide what to call, then the client executes the calls with whatever credentials the server holds.

02

The Main Risks

  • Tool poisoning: malicious instructions hidden in tool descriptions or results. See MCP tool poisoning.
  • Over-permissioned servers: servers holding broad API tokens or file access. See excessive agency.
  • Untrusted sources: community servers installed without review, or changing after installation.
  • Credential exposure: tokens stored in local configuration files.
  • Confused deputy: one server's output causing the agent to misuse another server's tools.
03

What to Test

  1. Inventory of every MCP server in use, its source and its permissions.
  2. Review of tool descriptions and update controls.
  3. Whether a malicious server in a test environment can influence other tools.
  4. Credential storage and scope.
  5. Logging of tool calls and approvals for sensitive actions.
04

How to Reduce the Risk

Maintain an allow-list of approved MCP servers, scope credentials tightly, pin and review tool definitions, require human confirmation for sensitive actions, run servers with least privilege and log all tool activity. Our AI agent and MCP security testing covers each of these.

05

Regulatory and Framework Context

For US organizations adopting AI agents, MCP governance belongs in the same supplier and access reviews used for other software and service accounts.

Being exploited now

Threat Radar →

The latest related vulnerabilities confirmed as exploited in the wild.

Latest intel on this topic

Updated hourly from the daily briefing.

Common questions

Who created MCP?

MCP was introduced by Anthropic in late 2024 as an open protocol and has since been adopted widely across AI tools and platforms.

Should we allow developers to install MCP servers?

Only from an approved list, with the same review you would give any software that holds credentials to company systems.

Change log

  1. November 13, 2026First published.
(877) 657-8496