// AI threat desk · 15 Nov 2026
Home/SUPPLY · AI Supply Chain
AI Supply ChainCriticalMCPCVSS 9.6

Corelight: MCP server setups let dependencies change unnoticed

Corelight said 23 of 35 reviewed MCP servers had recommended installs that would resolve dependencies fresh from a public registry. It published the findings on Nov 10, 2026.

AI-generated image of a security engineer reviewing server dependency files on blurred monitors in an operations centre.
AI-generated image, not a photo of the event.
Key takeaways
  • Corelight said one MCP server ran with 15 different sets of dependencies over five weeks without any configuration change.
  • Corelight said 23 of the 35 MCP servers it reviewed had a recommended install that would have resolved dependencies fresh from a public registry.
  • Dependencies run in the MCP server’s process and can access its API tokens, user files and reachable network paths.

Corelight published its report on Nov 10, 2026. It said one MCP server ran with 15 different sets of dependencies over five weeks without any configuration change.

Corelight said 23 of the 35 MCP servers it reviewed had a recommended install that would have resolved dependencies fresh from a public registry. Code in any dependency runs with whatever the server has, including API tokens, user files and network paths.

Corelight said the issue came from setup choices rather than MCP itself. Its review covered servers from two vendor catalogs and its own lab, and also found more than 1,600 connections in one week from MCP servers to 11 management systems.

On this page

uvx rebuilt one MCP server into 15 dependency sets

Corelight found one MCP server launched with uvx, a one-step tool that downloads a Python package and runs it from a single configuration line. The launcher resolved dependencies from PyPI whenever it rebuilt the environment, so 23 of 69 packages changed versions beneath a release that appeared pinned.

The project had pinned seven direct dependencies and published a hash-pinned lock file, but the launcher installed from the package’s declared dependencies instead of that lock file. Its cache held 19 builds of the same release and commit over five weeks, producing 15 distinct environments.

Another commit-pinned server produced 17 distinct environments across 19 builds. Its MCP framework was two major versions ahead of the version in the project’s own lock file.

23 of 35 reviewed MCP servers fetched dependencies afresh

Corelight reviewed 35 MCP servers from two vendor catalogs and its own lab. It found that 23 had recommended installations that would resolve dependencies fresh from a public registry.

In several vendors’ official setup instructions, local MCP proxies started with npm’s one-step runner and a latest tag. That caused clients to request the newest version from the npm registry each time they started.

One proxy released six versions in one recent month, including four within 14 hours. Corelight said its testing showed that freezing five servers kept their installed package lists unchanged across 457 starts over two weeks.

MCP dependencies can reach tokens and management systems

Corelight said code in any dependency, including a package no one deliberately selected, runs with whatever access the server has. That includes API tokens in its environment, the user’s files and network paths reachable from the workstation.

The same telemetry recorded more than 1,600 connections in one week to 11 management systems. Those systems included the firewall, the SIEM and sensor management.

Corelight cited a 2025 critical command-injection flaw in a widely used MCP proxy with a CVSS score of 9.6. It said a malicious or intercepted server could trigger that flaw on the user’s machine.

Corelight recommended pinning servers to commits, installing from hash-pinned lock files, launching installed servers directly and disabling startup update checks.

What to do now

  1. Pin to a commit, not a tag. Install from the project’s hash-pinned lock file. Launch the installed server directly, not through the one-step launcher. Turn off startup update checks.

FAQ

What did Corelight find about MCP server installations?

Corelight said 23 of the 35 MCP servers it reviewed had a recommended install that would have resolved dependencies fresh from a public registry.

How often was one MCP server rebuilt?

Corelight said one MCP server ran with 15 different sets of dependencies over five weeks, and its cache held 19 builds of the same release and commit.

Who is affected by the MCP dependency issue?

AI coding assistants running MCP servers can be affected when dependencies run with access to API tokens, user files and reachable network paths.

How should MCP servers be installed?

Corelight recommends pinning to a commit, installing from a hash-pinned lock file, launching the installed server directly and turning off startup update checks.

Is the issue a flaw in MCP itself?

Corelight said the issue came from choices in setup instructions or configuration and was not a flaw in MCP itself.

Sources

  1. Hackers Could Exploit MCP Server Supply Chain Weaknesses to Steal API Keys and Administrator Credentials, HKCERTPrimary
  2. The Supply-Chain Risk Hiding in Your MCP Servers | Corelight, CorelightPrimary
  3. Stolen AI API keys are paying for hackers’ autonomous scans, Help Net Security
How we checked this story
ClaimSourceStatus
Corelight found AI coding assistants running MCP servers with privileged credentials on an administrator’s workstation, although nothing there was compromised.CorelightConfirmed
Corelight found that one MCP server ran with 15 different dependency sets over five weeks without any configuration change.CorelightConfirmed
The launcher resolved dependencies from PyPI whenever it rebuilt the environment, causing 23 of 69 packages to change versions underneath an apparently pinned release.CorelightConfirmed
The MCP server was launched with uvx, which downloads a Python package and runs it from one configuration line.CorelightConfirmed
The project had pinned seven direct dependencies and published a hash-pinned lock file, but the launcher did not install from that lock file.CorelightConfirmed
The launcher’s cache held 19 builds of the same release and commit over five weeks, producing 15 distinct environments.CorelightConfirmed
One commit-pinned server had 19 builds and 17 distinct environments, while its MCP framework was two major versions ahead of its lock file.CorelightConfirmed
Corelight found that 23 of 35 reviewed MCP servers had recommended installations that resolved dependencies fresh from a public registry.CorelightConfirmed
Some vendors’ setup instructions start local MCP proxies with npm’s latest tag, causing clients to request the newest version whenever they start.CorelightConfirmed
A proxy released six versions in one recent month, including four within 14 hours.CorelightConfirmed
Corelight said the issue was caused by setup choices rather than a flaw in MCP itself.CorelightConfirmed
Dependencies run in the MCP server’s process and can access its API tokens, user files and reachable network paths.CorelightConfirmed
Corelight observed more than 1,600 connections in one week from MCP servers to 11 management systems.CorelightConfirmed
Corelight froze five MCP servers and found their installed package lists unchanged after 457 starts over two weeks.CorelightConfirmed
Corelight cited a 2025 critical command-injection flaw in an MCP proxy with a CVSS score of 9.6.CorelightConfirmed
Corelight recommends giving each MCP server its own least-privilege account or API client with expiring tokens.CorelightConfirmed
Bitsight found that mentions of AI agents, indirect prompt injection and MCP rose 67 percent across thousands of underground-forum posts from February through August 2026.Bitsight, via Help Net SecurityAttributed
Bitsight said the seller claimed 1,072 findings across 199 scanned hosts, but it could not verify that number.Bitsight, via Help Net SecurityAttributed
HKCERT published its report on Nov 11, 2026.HKCERTConfirmed
corelight.com published its report on Nov 10, 2026.corelight.comConfirmed

Could not verify

  • Whether any malicious dependency exploited the described MCP deployment weakness is not established
  • Whether any organization suffered credential theft or compromise is not established
  • Whether the CVSS 9.6 proxy flaw has a CVE identifier is not established
  • How many organizations use the affected MCP deployment patterns is not established
Explore with AI