A significant security vulnerability has been identified within the Model Context Protocol (MCP) ecosystem, where servers intended to run fixed, pinned versions of software are inadvertently pulling updated, unverified dependencies. Research conducted by Corelight revealed that an MCP server configured to run a specific release can undergo substantial dependency drift without any changes to its primary configuration. In one documented instance, a server maintained the same version number while cycling through 15 different sets of dependencies over a five-week period.
The issue stems from the way certain launchers, such as uvx, handle package installation. While administrators may specify an exact release tag, PyPI version, or Git commit, the underlying dependency resolution process often ignores hash-pinned lock files. Instead, the launcher may resolve the secondary packages directly from the Python Package Index (PyPI) each time the environment is rebuilt or initialized. Consequently, even when the primary MCP server code remains static, the dozens of supporting libraries—including critical components like cryptography and JSON Web Token (JWT) handlers—are subject to frequent, silent updates.
This behavior poses a severe security risk because MCP servers are frequently granted access to sensitive user credentials and internal management systems. Corelight observed a single workstation server making over 1,600 connections to 11 different management systems, including SIEM platforms, firewalls, and sensor management tools, within a single week. If a malicious actor compromises any of the secondary dependencies, they gain the same level of access as the MCP server itself. The danger is compounded by the frequency of these environment rebuilds; researchers noted that a single server was launched 162 times in one week, with the environment cache holding 19 different builds of the same version.
The prevalence of this issue is widespread across the industry. An analysis of 35 MCP servers from two vendor catalogs and internal lab environments found that 23 were configured to pull dependencies from public registries upon every launch or image build. Researchers noted that every administrator consulted during the investigation was utilizing at least one MCP server susceptible to this dependency drift. This pattern mirrors previous supply chain attacks, such as the June 2026 incident involving a Shai-Hulud worm variant that targeted AI coding assistant keys by compromising MCP-themed packages on PyPI.
To mitigate this risk, security teams are advised to move away from dynamic dependency resolution. Corelight demonstrated that by freezing servers—pinning them to a specific commit and installing them once from a hash-pinned lock file while disabling automatic update checks—the installed package list remains consistent over time. In a two-week test of five frozen servers, the package configurations remained identical across 457 launches, effectively eliminating the drift observed in non-frozen environments.
However, freezing code creates a secondary challenge regarding long-term maintenance and vulnerability management. When researchers froze a community server, they found that 40 of its packages were outdated, including some with known security sensitivities. The recommended approach is to balance immutability with proactive oversight. Administrators should perform nightly vulnerability reviews of their MCP servers and manually update dependencies only after verifying the changes, rather than allowing automated launchers to pull the latest versions from public registries.
The findings have prompted action among industry participants. Corelight confirmed that they have notified the maintainers of the 23 affected public MCP servers and the services that host them. As a direct result, at least one company hosting an AI agent exchange has overhauled its security review process to specifically evaluate submissions for this dependency-pinning vulnerability. Additionally, at least one vendor is currently in the process of completely repackaging their MCP server to address the flaw.
Ultimately, the incident highlights a critical blind spot in the deployment of AI-integrated tools. Any MCP entry initiated by a package runner that defaults to the latest available versions or fails to enforce strict hash-based locking is potentially vulnerable to code injection. As AI coding assistants continue to gain deeper integration into enterprise environments, the reliance on transparent and reproducible build processes remains a mandatory requirement for maintaining the security of the software supply chain.





