A Pinned MCP Server Is Not a Frozen One

November 16, 2026 · SPR{K3 Research

Corelight pulled telemetry from 35 Model Context Protocol servers and watched what they did on startup. Twenty-three of them reached out to a public registry every single time they launched. One of those servers was pinned to a specific commit. Over five weeks it ran fifteen different dependency sets anyway. The Cloud Security Alliance wrote it up in its November 15 CISO Daily Briefing under the title "MCP Servers: Pinned but Not Frozen."

If you treat "pinned" as a commitment that the code and its dependencies will be the same on Tuesday as they were on Monday, the finding is the news. Twenty-three out of thirty-five says most of the measured population; fifteen compositions in five weeks says the drift isn't a tail event. The commit hash in the manifest didn't lie — it still points at the same source tree. What changed sat one layer down, in the transitive dependency graph the MCP server pulls in on every start.

Why MCP makes this worse than a typical npm problem

An MCP server is not a library. It is a program that runs with the privileges of the agent that uses it — read your inbox, open shell commands, hit your internal APIs, write to your filesystem — and the agent reads its outputs as context. "Transitive dependency updated to a new version between runs" is a common thing in a lot of ecosystems, and it is mostly fine in most of them. In an MCP server it means the untrusted code with filesystem, network, and tool-call access just quietly rotated.

The named prior event is CVE-2025-6514 in mcp-remote, an RCE that let a malicious MCP server execute code on the host of the client that connected to it. The CSA briefing cites it as the kind of primitive this attack surface already fires through. Pin or no pin, if a dependency the server loads changes under you to something hostile, the game ends on the host side the moment the agent uses the server.

The cohort around this finding

This isn't an isolated data point. Three overlapping stories landed in the same two weeks:

Read together, those three say the same thing from three sides: the MCP server that your agent connected to this morning isn't guaranteed to be the MCP server your agent connects to this afternoon, even without a release.

What the CSA briefing recommends

The briefing's own guidance is straightforward and worth reproducing:

None of that is novel for traditional software supply chain hygiene. What is novel is the audience: AI engineering teams who shipped MCP servers quickly because the protocol was new and the integrations were useful, and who may not yet have the package-hygiene muscle of a mature platform team.

The defender's position

The useful mental model here is: the pin tells you what the operator intended; the runtime composition tells you what actually executed. Treat them as separate facts. The first lives in your manifest. The second has to be measured — at the host where the MCP server actually runs, at the time it actually runs — because nothing in the manifest alone will tell you a transitive dependency silently rotated.

Agent tool-call layers are a good place to put that measurement. The call is where the server's effects become observable: which binary got invoked, which network endpoints it reached, which filesystem regions it touched. If the fifteenth dependency set in five weeks suddenly tries to post to a domain it has never posted to, the pin doesn't catch that. Behavior does.

Corelight's finding doesn't say the MCP ecosystem is on fire. It says pinning, as most operators are doing it right now, is weaker evidence of safety than its name suggests. The reasonable response is to assume the runtime can change under you, measure what actually runs, and keep the trust boundary at the tool-call layer rather than at the version-string layer.


SPR{K3 is a security research operation that pairs offensive vulnerability research with runtime behavioral defense. Defend is our runtime agent. To talk about a deployment, reach us at support@sprk3.com.