
MCP in AI: how agents plug into your apps, and how a fake server stole emails
This week the FBI and Amazon described attackers handing whole stages of hacks to AI agents. The same agents need a way to reach your mailbox, files and code, and many of them now get it through MCP, the Model Context Protocol.
What MCP does
MCP is an open standard for connecting a model to outside tools and data. A developer writes one MCP server for a service, say an email provider or a database, and any app that speaks the protocol can use it. Without it, every chat app needed its own integration for every tool. An AI agent uses those servers to read, search and act on your behalf.
Anthropic open-sourced MCP in November 2024. On 9 December 2025 it donated the protocol to the Agentic AI Foundation, a fund under the Linux Foundation that Anthropic co-founded with Block and OpenAI. By then Anthropic counted more than 10,000 active public MCP servers and 97 million monthly downloads of its software kits, and ChatGPT, Cursor, Gemini, Microsoft Copilot and Visual Studio Code supported it.
“MCP started as an internal project to solve a problem our own teams were facing.”
— Mike Krieger, chief product officer at Anthropic
Quote source: Linux Foundation, 9 December 2025.
How a fake server stole emails
In September 2025, Koi Security found a package called postmark-mcp on npm, the main registry for JavaScript code. It copied the legitimate MCP server of the email service Postmark, and version 1.0.16 added one line that sent a hidden copy of every outgoing email to the attacker, The Register reported. The package had about 1,500 downloads a week. Postmark said the package was not theirs.
The attack needed no exploit. The server held a real API key, so the stolen copies looked like ordinary mail to the provider. Its users had given the server permission to send email, and the attacker used that permission.
Our take
Maks, the Intokened editor, works in Claude Code every day, and Claude Code, like Cursor and ChatGPT, can connect to MCP servers. That is where most people will meet the protocol: a button that offers to link a mailbox or a repository to the assistant.
Treat each MCP server like an app that gets your password. Check who published a package before you install it: Postmark did not publish the fake one. Install servers from the service's own repository, pin a version you checked, and give an API key only the rights the job needs, for example read-only access to a mailbox that the agent only summarises. Remove servers you stopped using, because each one keeps its key.
Our AI glossary explains the other terms around agents, from context windows to RAG.
The figures on MCP adoption come from Anthropic's announcement of 9 December 2025. The postmark-mcp details come from Koi Security's research as reported by The Register.

Author
MaksTrading man
I've been interested in the cryptocurrency market for a long time, am a trader, and write articles and news about my experience and crypto in simple terms.
Comments (0)
No comments yet — be the first!
Join our community
No spam, unsubscribe in one click
Related news

AI video generators compared: Runway, Kling, Luma, HeyGen and Synthesia prices

WhiteBIT turns eight: from crypto exchange to a global ecosystem of 40M users

RWA tokens ahead of tokenized stocks: Quant doubled in a year, Ondo fell 43%
Most readTop 7
Silicon Valley Workers Are Wearing Noise-Cancelling Masks to Dictate AI Prompts
398AI


