← All Labs
🛡 AI HACKING ADVANCED +135 XP · +54 no-hint bonus

Rogue MCP Server — Review Tools Before You Connect

Your company's AI agents get new powers by connecting to MCP tool servers. Six are waiting for approval. Tool descriptions are read by the model as instructions, scopes decide what a tool can touch, and servers can change after you approve them. Three of these six should never be connected.

https://bookshop.local/search

What is a rogue MCP server?

The Model Context Protocol, MCP, lets an AI agent connect to servers that give it new tools. Connecting a server is like installing software: its code runs with the access you grant, and its tool descriptions are read by your model.

A rogue MCP server is one that abuses that position. It may copy data passing through it, ask for more access than it needs, or carry hidden instructions in its descriptions.

A real case

postmark-mcp, a malicious MCP server on npmSeptember 2025

Koi Security found that an npm package imitating Postmark's MCP server worked normally for fifteen releases, then added one line that sent a hidden copy of every email to an address controlled by its author.

This lab puts you in the position those users were in: deciding which tool servers to trust before connecting them.

Source: The Hacker News

How to fix it

  • Review an MCP server's source, publisher and requested permissions before approving it.
  • Pin the version you reviewed and review again before upgrading.
  • Read the full tool descriptions the model will see, not only the names.
  • Run third-party servers in isolation with the least access they need.

Questions people ask

What is MCP?

The Model Context Protocol is an open standard for connecting AI assistants to tools and data sources through small servers.

How do I vet an MCP server?

Check who publishes it, read the code or have it reviewed, compare the access it asks for with what it claims to do, and pin a version.

Why is a version pin important?

A server can behave well for many releases and change later. A pin means an update cannot reach you without a new review.

Read the full OWASP LLM Top 10 guide · All free labs · AI red team certification