mythsthe AIpostmark-mcp npm package2025AI tooling / email

everyone assumed:

the email connector with the vendor's name on it came from the vendor

what was actually true:

The npm package called postmark-mcp was published by someone unconnected to Postmark, and from version 1.0.16 it quietly copied the emails it sent to an outside address.

how it happened

  1. whoan unrelated publishercopied Postmark's open code to npm
  2. the gapan assistant tool nobody vettedinstalled by name, updated without review
  3. thena hidden BCC on every sendcopies go to an outside address
  4. the resultsent mail copied outscale unknown, estimates disputed
the gap that let it happen the result

what everyone assumed

A connector called postmark-mcp, built for the Postmark email service, looks like Postmark’s own. The name matches, the code matched Postmark’s open repository line for line, and it did the job: an AI assistant could send email through it. When a tool works and carries the vendor’s name, checking the publisher feels like paperwork. Updates feel like maintenance.

what happened

MCP servers are small programs that let an AI assistant use another service. Postmark’s official one asks for a Postmark server token, per its own GitHub documentation, and then sends whatever the assistant tells it to send. In September 2025 Postmark kept that server on GitHub only. Its statement says it had never published the server on npm before the incident.

Someone else did. The npm registry’s own record shows a package named postmark-mcp created on 15 September 2025. Koi Security, the researchers who found it, say the publisher took the legitimate code from Postmark’s repository and published it under the same name. The registry shows version numbers 1.0.0 to 1.0.15 going up within about 26 hours, across 15 and 16 September. Postmark describes this as a malicious actor who “built trust over 15 versions”.

On 17 September 2025, at 08:59 UTC, version 1.0.16 went up. Koi says the only change was one added line that put a hidden BCC on every outgoing email, to an address at giftshop[.]club. Two more versions followed within half an hour, and Koi lists 1.0.16 and every later version as malicious. From then on, each email the assistant sent went to its real recipient and to the outside address. As Koi points out, the assistant had no way to notice: the tool still sent the email and reported success.

Postmark says its own API and services were never compromised. The package needed no break-in: it ran with whatever the user had handed it.

step by step

  1. 15 to 16 September 2025a package named postmark-mcp appears on npmThe npm registry shows versions 1.0.0 to 1.0.15 published within about 26 hours, built from Postmark's own open code.
  2. 17 September 2025version 1.0.16 adds a hidden BCCEvery email the tool sends now also goes to an address at giftshop[.]club, according to Koi Security and Postmark.the assumption breaks here
  3. 25 September 2025the package disappears from npmThe registry shows every version unpublished at 03:31 UTC; copies already installed keep running.
  4. 25 September 2025Koi Security and Postmark go publicPostmark says it never published the package and tells users to remove it and rotate credentials sent by email.

how it surfaced

Koi Security published its findings on 25 September 2025, the same day Postmark published a warning. Koi is a security vendor that sells tooling for exactly this kind of risk, so its numbers come with that interest attached. They are also, in Koi’s own words, a guess: “We can only guestimate the impact”.

  • Downloads. Koi says the package was downloaded about 1,500 times a week. npm no longer serves download counts for it, so the figure cannot be checked.
  • Use. Koi assumes about 20% of those were in active use, about 300 organisations, each sending 10 to 50 emails a day.
  • Emails. That gives Koi’s estimate of 3,000 to 15,000 emails a day copied out. It is a calculation from assumptions, with no count behind it.
  • Known customers. Postmark told The Register on 29 September 2025 that it knew of only one of its customers that used the package.

The npm registry shows every version unpublished on 25 September 2025 at 03:31 UTC. Koi says the developer deleted the package after Koi contacted him; CSO Online reported it was gone before Koi could report it to npm. Neither source changes the main point: removal from npm does nothing to copies already installed. Postmark told anyone who used it to remove it, check email logs for suspicious activity and consider rotating any credentials sent by email during that period. On 26 November 2025 Postmark’s official server appeared on npm under the scoped name @activecampaign/postmark-mcp, so the vendor’s own package now sits next to the name the impostor used.

The other MCP case on record put the attack in a tool’s description. This one needed nothing so clever: a familiar name and an update. tj-actions showed the same trust in a version number, inside a build pipeline.

A tool that carries a vendor’s name was written by whoever published it, and every update brings in new code.

The agent permissions guide covers what an assistant’s tools can do with your authority, and the mail guide covers why mail leaving your company is so hard to see from the admin side. The MCP servers guide goes through inventory and pinning in full.

picture it at a company like yours

The company here is made up. The other two paths are not: they happened, and each links to the story.

your company
a normal daySomeone on the team connects an AI assistant to the email service so it can send customer updates.
the gap nobody seesThey install the connector that carries the vendor's name, and nobody checks who published it or what the next update changes.trusting what a supplier ships
  1. in this story
    the day it mattersAn update adds one quiet line, and every message the assistant sends gets a hidden extra recipient.
    what it costs youInvoices, password resets and customer details sit in a stranger's inbox, and you only learn it from someone else's research.
  2. or it goes like this
    howA genuine, signed update from a certified vendor carries a backdoor.
    what leavesa quiet way into your networkwho ends up with ita state intelligence service
    it happenedSolarWinds / Sunburst, 2020
  3. or it goes like this
    howA build helper everyone uses is hijacked and its versions repointed.
    what leavessecret keys, printed into build logswho ends up with itanyone reading public logs
    it happenedtj-actions/changed-files, 2025

sources

  1. Postmark: Information regarding malicious "postmark-mcp" package, 25 September 2025
  2. Koi Security: First Malicious MCP in the Wild: The Postmark Backdoor That's Stealing Your Emails, 25 September 2025 (archived copy)
  3. npm registry: postmark-mcp package metadata (publish and unpublish times)
  4. npm registry: @activecampaign/postmark-mcp package metadata
  5. GitHub: ActiveCampaign/postmark-mcp, official Postmark MCP server
  6. The Register: One line of malicious npm code led to massive Postmark email heist, 29 September 2025
  7. CSO Online: Trust in MCP takes first in-the-wild hit via squatted Postmark connector, 26 September 2025

← all myths

your turn

what you can do about it

  1. this week

    List every MCP server on your team's machines and match each publisher to the vendor's own docs.

    Ask each person who uses an AI assistant to open its MCP configuration and send you the list of servers and packages it runs. For each one, find the vendor's documentation page for its official connector and compare the exact package name and publisher. Remove anything that does not match, pin the ones you keep to a version you have looked at, and rotate any token or password that went through a removed one. If a connector updated itself tomorrow, who would notice?

  2. go deeper

  3. if you're not sure where to start

    Run the offboarding checklist on one leaver. It covers every account and token to close when someone leaves.

    open the checklist →

the letter · every two weeks

one access gap in your inbox, with the fix.

Every issue takes one way access creeps in, shows a real incident where it went wrong, and ends with one thing to check that week. Written for the people who run the tools, without a security team behind them.

  • 1one mechanism, explained plainly
  • 2one real incident, taken apart
  • 3one step you can take on Monday

get the next issue

One email every two weeks. Your address stays here and is never passed to vendors. One click to unsubscribe. Privacy.