in this guide
In September 2025 a package called postmark-mcp on npm let AI assistants send email through Postmark. It worked well for fifteen versions. Version 1.0.16, according to Postmark’s own warning of 25 September 2025, “secretly BCC’d emails to an external server”. Postmark had never published the package. Its own MCP server lived on GitHub. Anyone whose assistant ran that version from npm was copying every message it sent to a stranger.
That is the mechanism in one story. An MCP server is a small program that sits between an AI assistant and one of your systems: email, the CRM, a database, the file share. Someone installs it so the assistant can do real work. From then on it runs on their laptop, holds a key to that system, and hands the assistant text that the assistant treats as instructions. Whoever publishes the next version decides what it does.
The agent guide covers the identity an agent acts under and how much that identity can reach. This guide is about the package in the middle: who wrote it, which version is running, what key it was given, and whether anyone at your company could list them. For the short explainer, see AI connectors and MCP.
what you hand over when you add one
Anthropic released the Model Context Protocol on 25 November 2024 as an open standard for connecting assistants to the systems where company data lives. Most assistants and coding tools now speak it. Adding a server to one of them is usually a few lines in a configuration file.
The MCP project’s own setup tutorial shows what those lines look like. The assistant is told to start the server with npx -y and a package name, with no version number. npm’s documentation says npx asks before installing a package, and -y answers that question in advance. When a server needs an API key, the tutorial puts it in an env block in the same file, in plain text. And the tutorial carries a warning worth reading twice: the server runs with the permissions of your own user account, so it can do to your files whatever you can.
what a local MCP server can reach
- the key in the config filean API key or token for the system it connects to, readable by anything that can read the file
- your user accountfiles, SSH keys, other config files and browser data on the same laptop
- the assistant's attentiontool names and descriptions the model reads and follows
- the next versionwhatever the publisher ships, started at the next launch when no version is pinned
granted the moment one line is added to a config file
None of the four doors shows up in your identity provider. There is no consent screen and no admin approval, and the key in the file was often created by the same person who pasted it there, with whatever scope the dashboard offered first.
three ways it has already gone wrong
A look-alike package turned on its users. The postmark-mcp case is the cleanest example. The Register reported on 29 September 2025 that the code appears to have been Postmark’s own GitHub code with one line added. Koi Security, which found it, counted about 1,500 downloads in a week, per The Register. Postmark told The Register it knew of one customer who had used it. The scale is uncertain. The pattern is clear: a name that looked official, a history of clean versions, then one update. The sibling incident page tells it in full: the Postmark lookalike.
The tool description carried the attack. On 1 April 2025 Invariant Labs, a vendor that sells AI security tooling, published what it called tool poisoning: instructions hidden in a tool’s description, invisible to the user and read by the model. In its demonstration on one coding assistant, a harmless-looking tool led the model to read another server’s config file and the user’s SSH keys and send them out. Invariant also described what it called a rug pull, where a server changes its tool descriptions after the user has approved it. The incident page is the tool description attack.
The official server leaked across customers. Asana launched its own hosted MCP server on 1 May 2025. On 4 June it found a bug that, in the words of its notice to customers reported by The Register, could have exposed information from one customer’s Asana domain to other customers using the MCP server. The server was offline from 5 to 17 June and every connection had to be set up again. The Register reported no indication that the bug was exploited. Nothing was installed from a stranger here. The point is narrower: a connector is a new code path into your data, and the vendor’s first version of it is new code.
The connector is software. Someone ships the next version, and it runs with what you gave the last one.
what the specification covers, and where it stops
The MCP specification has a full authorization chapter, and it is good on the parts it covers. It also says at the top that authorization is optional, and it draws a line most installs fall on the wrong side of.
two kinds of MCP server, two sets of rules
remote server over HTTP
covered by the spec's OAuth rules
- Tokens must be issued for that server and checked by it
- The server must not pass your token on to other services
- Scopes should start small and grow on request
- Revoked like any other OAuth grant
local server on a laptop
the common way to install one
- The spec says it should take credentials from its environment
- The key sits in a config file in plain text
- Runs with the user's own account permissions
- Whatever version the launch line fetches
For servers reached over HTTP, the 2026-07-28 version of the specification is strict. A server must only accept tokens issued for it, and must never forward a token it received to another service. The project’s security best practices list that forwarding, called token passthrough, as forbidden, and warn against granting every scope up front, because a broad token widens the damage when it leaks.
For local servers, which talk to the assistant over standard input and output, the same chapter says they should not follow the authorization rules at all and should “retrieve credentials from the environment”. That is a sensible engineering choice. It also means the protection for the most common install is whatever the person who set it up did with their key.
The security best practices do address local servers. They describe the attack plainly: a malicious startup command in a client’s configuration, or a malicious payload inside the server itself. They require any client that offers one-click setup to show the exact command, untruncated, and get explicit approval. They recommend sandboxing and a warning that servers run with the client’s privileges. Those are obligations on the people who build assistants. Nothing in them checks who published the package.
The MCP project’s own registry, launched in preview on 8 September 2025, does not close that gap either. Its launch post describes server details as self-reported, with problems handled by community reports and removal after the fact.
finding what is installed
Each assistant keeps its list of servers in its own file, which is why nobody has the list. Two documented examples:
- Claude Desktop keeps it in
claude_desktop_config.json, under~/Library/Application Support/Claude/on macOS and%APPDATA%\Claude\on Windows, per the MCP project’s tutorial. - Claude Code keeps personal servers in
~/.claude.jsonand shared project servers in a.mcp.jsonfile at the root of a code repository, per its documentation. Runningclaude mcp listprints them. The same documentation says it asks before using a project’s servers in an interactive session, and loads them without asking in non-interactive runs.
Other assistants and code editors use their own files; Invariant’s demonstration read one at ~/.cursor/mcp.json. The vendor’s documentation for each tool your team uses is the place to look.
The .mcp.json case deserves a second look. A server listed in a repository travels with the code. Everyone who clones the project gets the same launch line, and a change to that file is a change to what runs on every developer’s laptop, which makes it worth the same review as a change to the build. The engineer’s laptop guide explains why that machine already reaches more than anyone planned.
pinning: making each update a decision
The browser extensions guide makes the point that an extension is whatever its latest update says it is. A local MCP server launched from a line with no version number has the same property, and the postmark-mcp users found out at version 1.0.16. The npm worm on the Shai-Hulud page shows the other route to the same place: a trusted publisher’s account is taken and new versions go out under their name.
Pinning changes who decides. In practice:
- Write the exact version into the launch line (
some-server@1.4.2). A bare name or@latestlets the publisher choose; a pinned number means an update happens when someone edits the file. - Match the publisher to the vendor’s own documentation. If a service has an official MCP server, its docs say where it lives. A package with the right name from a different publisher is a stranger.
- Treat an update like a new install. Read what changed, then edit the version. The rug pull Invariant described is a reminder that the tool descriptions can change too, so a new version deserves a fresh look at what the server tells the model.
- Prefer one server per job. A server that can only read the CRM does less harm than one that can read and send email, whatever version it is on.
the key outlives the server
Removing a server from the config file stops it from starting. It does nothing to the API key that was pasted into the same file, which still works for anyone who copied it, including a malicious version that ran for a week. Postmark’s advice to affected users covered the same ground: remove the package, check email logs, and consider rotating any credentials sent by email while it ran.
So each server should get its own key, created for it, named after it and scoped to the one job: send-only for email, read-only for reports. When the server goes, the key is revoked the same day. The token guide explains why a credential keeps working after the thing that used it is gone.
When the next version of the server your sales team uses to send email comes out, who at your company will read what changed before it runs?
on the record
all myths →sources
- Anthropic: Introducing the Model Context Protocol (25 November 2024)
- Model Context Protocol specification, version 2026-07-28: Authorization
- Model Context Protocol: Security Best Practices (2026-07-28)
- Model Context Protocol docs: Connect to local MCP servers
- npm Docs: npx
- Claude Code docs: Connect Claude Code to tools via MCP
- Postmark: Information regarding the malicious postmark-mcp package (25 September 2025)
- The Register: Postmark MCP server code hijacked (29 September 2025)
- Invariant Labs: MCP security notification, tool poisoning attacks (1 April 2025)
- The Register: Asana MCP server bug (18 June 2025)
- Model Context Protocol blog: Introducing the MCP Registry (8 September 2025)
