Skip to content
AI Security Wire

Published

- 5 min read

By

Corelight: A "Pinned" MCP Server Ran 15 Dependency Sets in Five Weeks

img of Corelight: A "Pinned" MCP Server Ran 15 Dependency Sets in Five Weeks

Corelight says an MCP server configured to run one fixed release on an administrator’s workstation ran with 15 different sets of dependencies over five weeks, with its configuration untouched. The finding comes from Corelight’s own lab, reported by James Lagermann, the company’s Director of Technical Alliances, in a 10 November blog post and a 13 November Help Net Security article quoting him. The 15-environment and 162-start figures come from a single lab: one workstation and one community server. Twenty-three of the server’s 69 packages changed versions, including the cryptography library, the JSON Web Token library and the MCP software development kit itself.

What Corelight found

The workstation’s AI coding assistants started their MCP servers with uvx, a launcher that downloads a Python package and runs it from a single line of configuration. One entry, a community MCP server for a network gateway, named an exact release tag. The others named an exact PyPI version or an exact Git commit.

Corelight says the project’s author had pinned the seven direct dependencies and published a hash-pinned lock file. But the launcher installed from the package’s declared dependencies, not the lock file, so roughly 60 packages underneath were resolved fresh from PyPI within the allowed version ranges whenever the environment was rebuilt.

The assistant started the server 162 times in one week. Most launches reused a cached environment. The cache held 19 builds of the same release and commit, made over five weeks, about one every couple of days, and those produced 15 distinct environments. During Corelight’s test, an ordinary launch built a new environment containing two dependencies released since the previous day’s build. Corelight says every MCP server launched this way on that workstation showed the same pattern, whether pinned to a tag, a PyPI version or a commit.

The pattern is not limited to Python. Corelight says several vendors’ official setup instructions start a local proxy with npm’s runner and a latest tag, so the newest version is requested at each start. In one other vendor’s example, the setup also turned off TLS certificate checking for the whole proxy process.

How widespread it is

Of the 35 MCP servers Corelight reviewed, 23 had a recommended install that would resolve dependencies fresh from a public registry, either on every launch or on each install or image build. Lagermann told Help Net Security: “I have seen the same pattern in customer environments and in other real-world deployments. So far, 100% of administrators that I’ve talked to were using at least one MCP Server with this issue.” That figure comes from Lagermann’s conversations with administrators, as quoted by Help Net, and is not a measured rate.

He said he notified maintainers and/or services that share the 23 public servers; one company hosting an AI agent exchange has since updated its security review of submissions, and in another case a vendor is repackaging its MCP server.

Why the stakes are higher for MCP

An MCP server acts for a user, so it often holds that user’s credentials. On this workstation, the servers made more than 1,600 connections in a week to 11 management systems, including the firewall, the SIEM and sensor management. Corelight notes that every dependency loads into the server’s process and so runs with the server’s access, and that the assistant starts the servers automatically each session, so rebuilds happen without anyone running an install command.

Per Corelight, citing Endor Labs and SecurityWeek (which we did not open), a June 2026 variant of the Shai-Hulud worm reached PyPI, including MCP-themed packages, and went after AI coding assistant keys; in one wave, it says, six malicious releases were published in under a minute and quarantined within 30 minutes. Its argument is that a malicious release may be live for only minutes or hours, so exposure depends on whether anything rebuilds in that window. No evidence found as of 13 November in the Corelight post or the Help Net article of any such release reaching the servers it studied. That reasoning is Corelight’s, not an observed incident.

Finding and fixing it

Corelight describes using DNS and TLS server-name logs to find hosts contacting GitHub, PyPI and the npm registry, then EDR process lineage to tie the connections to the assistant starting the launcher. In its lab that narrowed seven days of lookups to four hosts, and only the administrator’s workstation was installing from PyPI as assistant sessions started. Corelight cautions that contacting a registry is not evidence of drift on its own, since a cached launcher may only be checking a tag.

Its fix is to freeze the environment, not just the release: pin to a commit rather than a tag, install from a hash-pinned lock file into a dedicated environment with hash checking enforced, launch the installed server directly, and turn off startup update checks. Corelight froze five servers this way, and over the next two weeks, across 457 starts, each server’s installed package list matched its install-day list.

Frozen code ages. When the team froze one community server, 40 of its packages were older than what the launcher would have pulled that day, some security-sensitive. Lagermann advised following the MCP maintainers’ update recommendations; he said the servers he maintains are reviewed nightly for vulnerabilities and updated the next morning if one is found.

Practical takeaways

  • Treat any MCP entry started by a Python or npm package runner, or tagged latest, as code that can change between sessions.
  • Corelight’s advice: prioritize servers that hold administrator or management-plane credentials.
  • Pair freezing with a scheduled review so pins do not become stale.

References

Frequently Asked Questions

Is this a vulnerability in MCP?
No. Corelight's James Lagermann writes that none of this is a flaw in MCP and that MCP servers do not have to run this way. Each case was a choice in the setup instructions or the configuration, such as launching a server with the uvx runner from its declared dependencies instead of from the project's hash-pinned lock file.
Was any MCP server actually compromised?
Corelight says nothing on the workstation was compromised. The finding is that the launch method let the code change without anyone asking, so, Corelight argues, a hijacked dependency release could have been picked up on the next rebuild. No evidence found as of 13 November in the Corelight post or the Help Net article that a malicious package reached these servers.
How common is the pattern in what Corelight reviewed?
Corelight reviewed 35 MCP servers from two vendor catalogs and its own lab. Twenty-three had a recommended install that would resolve dependencies fresh from a public registry, either on every launch or on each install or image build.
What does Corelight recommend?
Freeze the environment once rather than only pinning the release: pin to a commit instead of a tag, install from the project's hash-pinned lock file into a dedicated environment with hash checking enforced, launch the installed server directly instead of through the one-step launcher, and turn off startup update checks. Corelight notes that frozen code ages and needs a deliberate update routine.