system: OPERATIONAL
← back to all hacks
SUPPLY CHAIN MEDIUM NEW

MCP Server Dependency Drift: One Pinned Release, 15 Different Environments

Corelight's November 2026 study shows MCP servers started by AI assistants quietly rebuilding from PyPI, running unreviewed code next to administrator credentials. How to freeze them.

2026-11-12 // 6 min affects: mcp-servers, ai-coding-assistants, uvx-launched-servers

What is this?

MCP server dependency drift is the gap between the version of an MCP server you think you run and the code that actually executes. On November 10, 2026, Corelight published “The supply-chain risk hiding in your MCP servers”, by James Lagermann, after a blast-radius review of an administrator’s workstation where AI coding assistants were launching several local MCP servers. GBHackers covered the research on November 11, 2026.

The core finding fits in one sentence from the post: pinning a package’s version does not necessarily pin the packages it depends on. A single MCP server, configured to one release and one commit, ran with 15 different dependency sets over five weeks. Those servers held administrator credentials for the firewall, the SIEM and sensor management.

How it works

The servers were started through a one-step launcher, uvx, as recommended by their install instructions. The uv documentation describes the environment created this way as stored in the cache and “treated as disposable”: if it is removed, a new one is created automatically. Each rebuild resolves dependencies again from the public registry.

Corelight measured what that means in practice:

ObservationFigure
Builds of the same release and commit found in the launcher cache19
Distinct environments among those builds15
Packages that changed version (out of 69)23, including the cryptography library, a JWT library and the MCP SDK
Launches of one server by one assistant in one week162
MCP servers reviewed whose recommended install resolves fresh from a public registry23 of 35

The maintainer of the drifting server had done the right things on paper: seven direct dependencies pinned and a hash-pinned lock file published. The launcher never used the lock file, so roughly 60 transitive packages were resolved fresh at each rebuild. Every server launched this way on the workstation drifted, whether it was referenced by tag, by PyPI version or by commit.

The network view confirmed the exposure. Using Zeek DNS logs, TLS server names and EDR telemetry, Corelight attributed 203 GitHub connections and 136 PyPI connections in one week to the workstation, the only host installing from PyPI, started by two AI assistants. In the same week, the MCP servers made more than 1,600 connections to 11 management systems.

Corelight frames this as a setup choice, not a flaw in the MCP protocol. No malicious package was found.

Why it matters

A drifting MCP server is a standing invitation to the next registry compromise. The post recalls a critical command-injection flaw (CVSS 9.6) in a widely used MCP proxy, disclosed by JFrog in July 2025: the proxy mcp-remote allowed a malicious MCP server to run OS commands on the client, in versions 0.0.5 to 0.1.15, fixed in 0.1.16. Corelight also notes that a proxy used by hosted MCP servers shipped six releases in one recent month, four of them within 14 hours. With an unpinned launcher, each of those releases reaches the workstation without anyone reviewing it.

The location makes it worse. These processes run on the machines that hold the most privileged credentials, and they rebuild silently, so the usual software inventory shows one version while the process runs another. Worms such as Shai-Hulud, which reached at least 187 packages in its first days and targeted AI assistant keys, show how fast a poisoned dependency propagates. This complements our coverage of agent skills as a supply chain, the tool update gap and the LiteLLM PyPI startup-hook stealer.

Defenses

Corelight’s recommendations, which it tested on five servers started 457 times over two weeks with package lists identical to the install-day record:

  • Pin to a commit, not a tag. Tags can move; commits cannot.
  • Install from a hash-pinned lock file with hash checking enforced, into a dedicated environment per server.
  • Launch the installed server directly, not through a one-step launcher that may rebuild it.
  • Turn off startup update checks in the server and its proxy.
  • Scope credentials. Give each server read-only, short-lived credentials limited to what its tools need.
  • Name an owner. Freezing is not a one-off: one frozen server already had 40 packages older than what the launcher would have fetched that day. Someone must re-lock, review and scan on a schedule.

A side benefit: startup fell from 4 to 11 seconds to about 2 seconds once rebuilds stopped.

Additional checks for defenders: inventory every MCP server configured in each assistant on admin workstations; watch for package-registry traffic from hosts that should not install software; and count cache directories to spot servers that rebuild repeatedly. For maintainers, Corelight asks for hash-pinned requirements, bounded version ranges, pinned proxies and documented startup network behavior.

Status

ItemDetail
ResearchCorelight blog, published 2026-11-10
NatureConfiguration and supply-chain risk; no malicious package found, no new CVE
Root causeOne-step launchers re-resolving transitive dependencies from public registries
Related flaw citedmcp-remote 0.0.5 to 0.1.15, fixed in 0.1.16 (CVE-2025-6514, CVSS 9.6)
MitigationCommit pin, hash-pinned lock file, direct launch, no startup updates, scoped credentials

Sources