You pinned the version. The config has not changed in five weeks. Every session starts the same release tag.
Then you look at what actually ran.
One MCP server in Corelight's lab ran with 15 different sets of dependencies across five weeks, and nobody touched the configuration. Twenty-three of its 69 packages changed version, including the cryptography library, the JSON Web Token library and the MCP software development kit itself. The version number stayed the same the whole time.
Corelight's James Lagermann published the finding on November 13, 2026, and it lands on a question every team running agents now has to answer: which code is my assistant actually executing?
Primary source: Your pinned MCP server may not be as pinned as you think (opens in a new tab).
What the lab measured
MCP servers are the bridge that lets an AI coding assistant act on other tools, so they hold that user's credentials. On the Corelight workstation the servers made more than 1,600 connections in a week to 11 management systems, among them the firewall, the SIEM and sensor management. Code that arrives inside any of their dependencies runs with that access.
| Layer | What the configuration named | What the machine resolved |
|---|---|---|
| Launcher | One exact release tag, PyPI version or Git commit | uvx installed from the package's declared dependencies |
| Lock file | Present in the project | Skipped for one community server, so roughly 60 packages resolved fresh from PyPI on each rebuild |
| Package versions | Fixed at the top of the tree | 23 of 69 packages changed over five weeks |
| Launch pattern | One server, one release | 162 starts in a week, with 19 cached builds of the same commit |
Corelight then widened the check. Across 35 MCP servers drawn from two vendor catalogs and its own lab, 23 shipped a recommended install that pulls dependencies from a public registry on every launch, install or image build. Asked how often he sees that in the field, Lagermann said: "So far, 100% of administrators that I've talked to were using at least one MCP Server with this issue."
The timing window is what makes it sharp. A hijacked release can be live for minutes or hours, and a launcher that rebuilds its environment every couple of days will collect it. In June 2026 a Shai-Hulud worm variant on PyPI went after MCP-themed packages and the keys held by AI coding assistants.
The pattern: tool access is a supply chain
Every agent stack has two layers. The layer you review is a config file with a version string in it. The layer that runs is a dependency tree that resolves at start time.
assistant session
|
v
[ uvx / npx launcher ] resolves on every start
|
v
[ public registry ] code can change under you
|
v
[ MCP server process ] holds the user's credentials
|
v
[ firewall ] [ SIEM ] [ sensor management ] ... 11 systems
A version pin fixes the package you named. Its dependencies resolve on their own schedule, from a registry outside your control, into a process that carries credentials for the systems around it.
Freezing works, and it needs a cadence
Corelight froze five servers: pinned to a commit, installed once from a hash-pinned lock file, launched directly, update checks off. Over the next two weeks those five were started 457 times and every one's installed package list matched the list recorded on install day.
Frozen code ages, so the re-lock cadence carries the weight. One frozen community server held 40 packages older than what the launcher would have pulled that day, some of them security-sensitive. Lagermann's rule for the servers he maintains is a nightly vulnerability review, an update the next morning and a changelog entry to go with it.
How CommandLyne handles this work
The Corelight finding is a governance problem before it is a packaging problem. Someone has to know which tools an agent may call, which credentials those tools carry, and what the agent touched afterwards.
CommandLyne runs agents against a governed tool policy. Each user and each agent gets scoped tool access, so a server holds only the permissions its work needs. Agent execution runs inside its own VM-level isolation, personal and team memory is permission-aware, and every tool call lands in a full audit trail. Where a step writes to a system of record, the workflow stops at a human sign-off checkpoint and the run can be rolled back. Model calls route across Gemini, Claude, OpenAI and partner models under your own keys, with token-level cost visibility per task.
Start with the tool policy and MCP servers guide (opens in a new tab) for the scoping model, and Governance and Multi-Model Control (opens in a new tab) for how routing and cost reporting fit together. Security and Enterprise Memory (opens in a new tab) covers isolation, audit and customer-managed encryption.
Matching the control to the risk
| Question | Ad hoc MCP setup | Governed tool policy |
|---|---|---|
| Which code ran? | Whatever the registry resolved at start time | A frozen build, recorded at install and re-checked on a cadence |
| Who can call the server? | Anyone holding the config | Scoped per user and per agent |
| What can it reach? | Whatever the host network allows | An allowlist of destinations, with every call logged |
| Who approves a write? | Nobody, or a message in chat | A sign-off checkpoint with rollback |
| What did the run cost? | Invisible | Token-level cost per task |
Where to start
- Freeze the install. Pin the commit, install once from a hash-pinned lock file, launch the binary directly and switch update checks off.
- Set a re-lock cadence. A nightly vulnerability review, an update the next morning, a changelog entry alongside it.
- Give the server its own identity. A dedicated scoped credential beats a personal token that reaches 11 systems.
- Log every call. Session history and tool traffic belong next to your SIEM rather than in a local cache.
- Put a checkpoint on writes. Read freely, and require sign-off wherever an action changes a system of record.
Try CommandLyne to run agents against your own tools under scoped access, an audit trail and human sign-off, or book a demo with our team.
Try CommandLyne (opens in a new tab) | Book a demo (opens in a new tab)

