What happened
Corelight's November 2026 study ('The supply-chain risk hiding in your MCP servers') found that a single MCP server configured to one release and one commit ran 15 distinct dependency sets over five weeks because its uvx one-step launcher ignored the published lock file and re-resolved ~60 transitive packages from PyPI on each rebuild (23 of 69 packages changed version, including the cryptography library and MCP SDK). 23 of 35 reviewed MCP servers had a recommended install that pulls fresh dependencies from a public registry; one assistant's servers made over 1,600 connections to 11 management systems in a week. No malicious package was found in the review, but the setup choice is a standing supply-chain exposure, and the post documents prior registry-compromise precedent (xz, Ultralytics, Nx, Shai-Hulud worm, mcp-remote CVE-2025-6514).
Why it matters
MCP servers execute with the installing user's privileges and routinely hold the administrator credentials of the very security tools (firewalls, SIEMs, cloud consoles) they connect to an AI assistant. Because the launcher silently re-resolves transitive dependencies at every rebuild, a poisoned dependency lands in the same process as those credentials — turning MCP servers into a reachable supply-chain attack surface for AI deployments.
Attack vector
An attacker who compromises or typosquats any transitive dependency in an MCP server's unpinned dependency graph gains the code-execution context of the server process, which runs with the installing user's privileges and holds the credentials scoped to the connected security tooling/cloud consoles — enabling credential theft and lateral movement.
Affected systems
MCP servers launched via uvx/npx one-step launchers (Python/npm ecosystems), including self-hosted MCP servers holding privileged credentials (no single product)
Mitigation
Pin to a commit (not a tag), install from a hash-pinned lock file with hash checking into a dedicated environment per server, launch the installed server directly instead of through a re-resolving launcher, disable startup update checks, and scope each server's credentials to the minimum needed. Freeze and audit the environment on a schedule.