Get the free coreGet core

MCP tool poisoning and malicious MCP servers explained

Updated 6 min readSection 05 / Threats
SHORT ANSWER

MCP tool poisoning is an attack where instructions are hidden inside the description of an MCP tool, invisible to you but visible to the AI model. A malicious MCP server is a wider problem: the code itself works against you, as when the npm package postmark-mcp was reportedly modified to send email contents to an outside domain through a hidden copy (BCC). If you used it, uninstall it, rotate the credentials that went through it, and check your email logs. Background: what is an MCP server.

01MCP tool poisoning in plain words

An MCP server offers tools to your agent, and every tool has a human-readable description of what it does. You usually see a simplified version of it. The model sees the whole description, including hidden instructions.

Tool poisoning is the abuse of that gap: malicious instructions sit in the description, where you do not look and the model does. It is called a specialized form of indirect prompt injection. Prompt injection is a technique that tries to override or manipulate an assistant's instructions by inserting malicious text, and "indirect" means the text does not come from you but from something the agent reads. More on that: indirect prompt injection.

02What a poisoned tool can make an agent do

In the researchers' experiment, poisoned tools could instruct the model to access sensitive files such as SSH keys, configuration files and databases. SSH keys are the files that prove to servers and GitHub that you are you.

Why does the model obey? Prompt injections do not need to be readable by humans, as long as the model parses the content. Models will follow any instructions that reach them, whether or not they came from the operator or from some other source. A tool description is exactly such a channel, and it arrives every time the agent asks a server for its tool list. That is how MCP tool poisoning works without a single suspicious line on your screen.

A second finding matters if you connect several tools at once. A malicious server could do more than steal data: it could hijack the agent's behavior and override the instructions that other, trusted servers had provided.

03Four kinds of bad MCP server

People use "malicious MCP server" for several different things, and each needs a different defense. Only the first row below is tool poisoning proper; the other three are bad code and bad configuration.

KindWhere the bad part sitsWhat it needs from youCase
Tool poisoningText in a tool descriptionThe tool is connected and the model reads its descriptionHidden orders to read SSH keys
Malicious codeThe program itselfYou start it: a local server runs on your machine with your agent's privilegesAny local server you start
A good server turns badAn update to a server people already trustedNothing: you had already approved itpostmark-mcp, below
Bad configurationA start command placed in an agent's configurationYou accept a config you did not readCursor 1.2.4 and below: an approved harmless server replaced by a malicious command

For the first row the defense is fewer tools and tools from authors you can name, because you cannot read a description you cannot see. For the second it is vetting the author. The third needs a habit of looking again when something updates, and the fourth is accepting a configuration only after reading the exact command. The checklist is on MCP security.

04The postmark-mcp case: dates and versions

postmark-mcp is an npm package, an MCP server meant to let AI assistants send email through Postmark. It is a public example of a server that turned bad after people trusted it. Here is the timeline.

DateWhat happened
2025-09-15Version 1.0.0 was first published
2025-09-15 to 2025-09-17Rapid iteration through several minor versions. The behavior began around version 1.0.16 and persisted in later versions. Community reports allege that 1.0.16 introduced a hidden BCC that forwards every outbound email to an external domain
2025-09-17The latest noted release, 1.0.18, appeared on npm
2025-09-25The package no longer existed on npm and had likely been removed on detection

A BCC is a hidden recipient: the person you write to does not see it, and neither do you in the finished message. Exfiltrate means to send data out of your machine without your knowledge.

Why the data matters: MCP servers typically run with high trust and broad permissions inside agent toolchains, and what an email tool handles can be sensitive, such as password resets, invoices and customer communications.

This is not tool poisoning in the narrow sense, because nothing in a description had to be hidden. The server did what a mail tool does, plus one extra recipient. That is why it sits next to tool poisoning: the thing you approve is a name and a description, and the behavior behind it can change. The version numbers give you something to check in an old install, and an update later can change a server again, so read the launch command when it does.

05How to reduce the chance

None of this makes a server safe by itself, but each step cuts the damage a bad one can do.

06If you think you ran a bad server

For postmark-mcp the advice is: uninstall it, rotate any credentials ever sent through it, and review email logs for BCC traffic to the reported domain. Rotating means creating new keys and passwords and retiring the old ones, and it is safest from a device you trust.

To find out what is connected in Claude Code, run claude mcp list or claude mcp get <name> in your shell, or /mcp inside Claude Code. To remove a remote one, run claude mcp logout <name> first, because removing it alone does not revoke its sign-in tokens; logout revokes them where the service supports it. The full command reference is on Claude Code MCP. For keys, see leaked API keys, and if the server came from an npm package, see npm malware.

07MCP tool poisoning in short

Tool poisoning hides orders in the text of a tool description, where you do not look and the model does. A malicious server goes further and ships bad behavior in code or in a later update, as postmark-mcp did. Both ride on the trust you gave at the start, so vet the author, keep the list short, and go through the MCP security checklist before each new connection.

A tool description never shows up in a program list, so the part you can still watch is what happens on your Mac afterwards. A poisoned tool gives itself away only by what it makes the agent run, and the free, open-source core of Artificial Immunity watches MCP servers and every program start on your Mac and reports what it finds.

FAQ

What are the risks of using MCP?

A local server runs with the same privileges as your agent, its tool descriptions are text the model reads and you may not see, and Anthropic does not security-audit or manage any MCP server. The checklist is on the MCP security page.

How many MCP tools is too much?

No fixed number exists, but each tool adds a description the model reads, and each server is code or a service you trust. Keep only the ones you use.

What does MCP expose?

A server exposes capabilities to the agent. Each tool comes with a description the model reads, and in a tool poisoning attack that description carries hidden instructions.

Can you give me an example of AI poisoning?

In a researchers' experiment, instructions hidden in a tool description told the model to read SSH keys and configuration files. In a real case, the postmark-mcp package was reportedly modified to add a hidden copy (BCC) of email contents to an outside domain.

What is a malicious MCP server?

A server whose code or text works against you. It can run commands with your agent's privileges, carry poisoned tool descriptions, or change after you started to trust it.

Is tool poisoning the same as prompt injection?

It is a specialized form of indirect prompt injection. Prompt injection is a technique that tries to override or manipulate an assistant's instructions by inserting malicious text; in tool poisoning that text sits in the tool description.

Sources

  1. Invariant Labs: tool poisoning attacks
  2. Snyk: malicious MCP server on npm (postmark-mcp)
  3. MCP specification: security best practices
  4. Claude Code docs: MCP
  5. Claude Code docs: security
  6. NVD: CVE-2025-54136
  7. Artificial Immunity
A poisoned tool talks to your agent, not to you. Artificial Immunity watches the Mac where your agents work.Get the free core