---
title: What can real MCP security incidents teach enterprises?
url: https://opencontroller.lyzr.com/blog/lessons-from-real-mcp-security-incidents
published: 2026-11-17
updated: 2026-11-17
author: Lyzr
---

# What can real MCP security incidents teach enterprises?

**Real MCP incidents so far are supply-chain and local-tooling problems, not exotic AI attacks. A look-alike npm MCP server copied every email it sent to an outside address; a flaw in the `mcp-remote` proxy let a malicious server run commands on the user's machine; and the MCP Inspector developer tool accepted unauthenticated commands from a web page. Each points to a concrete control: pin and verify packages, inventory local MCP servers, and keep developer tooling patched.**

Below are three documented cases from 2025, what each one teaches, and a checklist you can apply today.

## Incident 1: what happened with the malicious postmark-mcp package?

In September 2025, researchers at Koi Security reported an npm package called `postmark-mcp`. It copied the name of an official Postmark Labs MCP server whose real source lives on GitHub. Early versions were a clean copy. Version 1.0.16, published on 17 September 2025, added a single line that BCC'd every email sent through the server to an attacker-controlled domain. Postmark stated it had no involvement with the npm package. The package had about 1,643 total downloads before the publisher deleted it ([The Hacker News](https://thehackernews.com/2025/09/first-malicious-mcp-server-found.html)).

**Lessons:**

1. **Trust built over 15 versions can be spent in one.** Approving a server once is not enough; updates need review.
2. **Name is not provenance.** The package name matched the real project; the publisher did not.
3. **An MCP server sees everything it handles.** An email-sending tool reads every message body and recipient.

## Incident 2: what was the mcp-remote vulnerability (CVE-2025-6514)?

`mcp-remote` is a small proxy that lets hosts which only speak local stdio connect to remote MCP servers. JFrog researchers found that versions 0.0.5 to 0.1.15 could be made to run operating-system commands when connecting to a malicious or hijacked MCP server, or through a man-in-the-middle on unencrypted HTTP. It was rated CVSS 9.6 and fixed in 0.1.16 ([GovInfoSecurity](https://www.govinfosecurity.com/serious-flaws-patched-in-model-context-protocol-tools-a-28924)).

**Lessons:**

1. **Connecting to an untrusted server can be enough.** The user did not need to call any tool.
2. **Glue packages are part of your attack surface.** Few inventories list proxies like `mcp-remote`, yet they run on developer laptops with full user rights.
3. **Plain HTTP to an MCP server is a real risk,** not a theoretical one.

## Incident 3: what was the MCP Inspector flaw (CVE-2025-49596)?

MCP Inspector is the official tool developers use to test MCP servers. Versions before 0.14.1 ran a local proxy that executed commands from any client without checking tokens or origins. A malicious web page could use cross-site requests and DNS rebinding to make the proxy start an arbitrary process with the developer's privileges. It was rated CVSS 9.4 and fixed in 0.14.1, which added a session token, localhost binding and origin checks ([Wiz vulnerability database](https://www.wiz.io/vulnerability-database/cve/cve-2025-49596)).

**Lessons:**

1. **Developer tools count.** The vulnerable component was a debugging tool, not a production server.
2. **"Localhost only" is not a security boundary** when a browser on the same machine can reach it.
3. **Global installs linger.** The fix applies only where someone actually upgrades, including project-local `node_modules`.

## Which controls map to which incident?

| Control | postmark-mcp | mcp-remote | MCP Inspector |
|---|---|---|---|
| Inventory of local MCP servers and packages on devices | Yes | Yes | Yes |
| Allow-list by verified publisher, not by name | Yes | — | — |
| Pin versions; review every update before use | Yes | Yes | — |
| Vulnerability scanning of MCP-related packages | — | Yes | Yes |
| Remote servers over HTTPS only, from an allow-list | — | Yes | — |
| Egress monitoring for unexpected destinations | Yes | Yes | Yes |
| Least-privilege credentials for each server | Yes | — | — |

## What should you do this quarter? A 10-point checklist

1. **Find every MCP server** on developer devices and in agent configs, including stdio packages and proxies. See [how to find local MCP servers on developer laptops](/blog/find-local-mcp-servers-on-developer-laptops).
2. **Record publisher and source repository** for each server, and confirm they match.
3. **Keep an internal allow-list**; block servers that are not on it. See [should you allow MCP servers from public registries](/blog/should-you-allow-mcp-servers-from-public-registries).
4. **Pin exact versions** in configs instead of `latest` or unpinned `npx -y` commands.
5. **Diff updates before approving them**, including tool descriptions. See [reviewing MCP server updates](/blog/review-mcp-server-updates-before-use).
6. **Track CVEs** for MCP SDKs, proxies and developer tools, and set a patch deadline (for example 7 days for critical).
7. **Require HTTPS** for every remote MCP server and reject plain HTTP.
8. **Give each server its own narrowly scoped credential,** never a personal admin token.
9. **Watch outbound traffic** from MCP processes for new domains.
10. **Have a removal playbook:** how to pull a server from every device in hours, and which credentials to rotate.

The wider governance model is in the [MCP governance pillar](/blog/how-to-govern-mcp-servers).

## How Opencontroller does this

Opencontroller Edge extends control to developer devices, coding agents, skills and local MCP servers, and is installed by IT through MDM. Discovered servers land in a governed registry of agents, versions, tools and MCP servers. Posture detects risky permission and configuration patterns, and policies as code can block servers that are not approved. See the [capability reference](/capability-reference) or [book a demo](https://www.lyzr.ai/book-demo/).

## Frequently asked questions

**Was the postmark-mcp package hosted in an official MCP registry?**
It was published on npm, the general JavaScript package registry. The lesson applies to any registry: check the publisher and the source repository, not just the package name.

**Are these vulnerabilities fixed?**
Yes. mcp-remote was fixed in 0.1.16 and MCP Inspector in 0.14.1. The remaining risk is old versions still installed on devices, which is why inventory matters.

**Do remote MCP servers avoid these problems?**
Partly. Remote servers remove local package installs, but you still need to verify the publisher, use HTTPS, scope OAuth tokens and watch for changed tool descriptions.

## Key facts

Documented MCP incidents: postmark-mcp npm package version 1.0.16 (17 Sept 2025) BCC'd every email to an attacker-controlled domain, reported by Koi Security, about 1,643 total downloads. CVE-2025-6514 in mcp-remote 0.0.5–0.1.15, CVSS 9.6, found by JFrog, fixed in 0.1.16, allows OS command execution when connecting to a malicious server. CVE-2025-49596 in MCP Inspector before 0.14.1, CVSS 9.4, missing authentication on local proxy exploitable via CSRF and DNS rebinding. Opencontroller Edge discovers local MCP servers on developer devices and is installed via MDM. Licensing starts from $250K per deployment per year.

## Related

- [Should you allow MCP servers from public registries?](https://opencontroller.lyzr.com/blog/should-you-allow-mcp-servers-from-public-registries.md)
- [How do you review MCP server updates before agents use them?](https://opencontroller.lyzr.com/blog/review-mcp-server-updates-before-use.md)
- [How do you find local MCP servers on developer laptops?](https://opencontroller.lyzr.com/blog/find-local-mcp-servers-on-developer-laptops.md)
- [More on Tools and MCP](https://opencontroller.lyzr.com/blog/cluster/tools-and-mcp)
