Skip to content
CSA Labs Research·Indexed by this site · Originally published· Tools →· Its place in the scan record →SelectedAttention39

Pinned but not frozen: dependency drift in coding assistant MCP servers

Pinned but Not Frozen: Dependency Drift in Coding-Assistant MCP Servers

AI summary

CSA published a research note pointing out that for local MCP servers launched by coding assistants via npx -y or uvx, even if the top-level version is pinned, transitive dependencies are still re-resolved according to semver ranges when the environment is rebuilt; the authors call this dependency drift. The note cites Corelight's review of 35 MCP servers: 23 of them resolve dependencies from public registries on recommended installation, and one server pinned by commit ran with 15 different dependency combinations within five weeks. The note uses the March 2026 axios supply chain incident as an example, in which a malicious version delivered a remote access trojan via an injected transitive dependency and was reportedly removed after about three hours; the postmark-mcp malicious server and CVE-2025-6514 in mcp-remote show that pinning can also freeze vulnerabilities. It recommends pinning exactly and inventorying servers, setting a release-age cooldown period, disabling install scripts, switching to locked builds or container images referenced by digest, and pairing these with a controlled update process.

Why it was selected

The original source explains clearly why pinning versions cannot lock transitive dependencies, and offers practices such as a cooldown period, disabling install scripts, and locked builds; teams can use it to check their own MCP configurations.

Read originallabs.cloudsecurityalliance.org