Table of Content
Pinned on Paper, Drifting on the Laptop: What Your MCP Config Doesn't Tell You
-
November 14, 2026
-
Security reviews approve a configuration. Agents run whatever that configuration resolves to on the day they start, and that can be different every time.
On 13 November 2026, Help Net Security reported on research by Corelight into MCP servers running on an administrator’s workstation in its own lab. One server was configured to run a single fixed release. Over five weeks, with that configuration untouched, it ran with 15 different sets of dependencies. 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 numbers around it are what make this a security problem rather than a tidiness problem:
- The AI assistant started one server 162 times in a week, and each start was a chance to pull new code.
- The workstation's MCP servers made more than 1,600 connections in a week to 11 management systems, including the firewall, the SIEM and sensor management.
- MCP servers often hold the user's credentials, so code that slips into any dependency runs with that access.
Pinning a package’s version doesn’t necessarily pin the packages it depends on. Corelight’s fix was to freeze five servers: pinned to a commit, installed once from a hash-pinned lock file, launched directly, with update checks turned off.
The same week, the tokens became the target
The day before, Help Net Security covered wfx-loot, a credential stealer that a threat actor pushed into tens of thousands of GitHub repositories from 10 November. It collects Actions secrets, cloud and SSH credential files, shell histories, and tokens for AI tools including OpenAI, Anthropic, Cursor and Copilot.
And on 12 November, Microsoft’s Defender Research and AI Red Team published “Trained to betray,” showing that poisoned fine-tuning data and backdoored model checkpoints could make coding agents issue unsafe tool calls on ordinary triggers. Their conclusion: treat post-training artifacts as part of the software supply chain, and enforce controls at the tool-execution boundary.
Put those together and the picture is clear. The AI toolchain on a developer or admin laptop is a supply chain that refreshes itself, and the credentials it holds are now a named target.
Why config review can't see this
Most MCP governance today checks what is declared: which servers are approved, which versions are listed in a config file, which tools they expose. That is necessary, and it is a snapshot. Drift happens at launch time, on the device, and it leaves no trace in the file a reviewer looks at.
Questions a config review can’t answer:
- What code actually ran? The resolved packages and their hashes at each launch, not the version string in the config.
- What did it load and read? Credential files, environment variables, SSH keys, AI tool tokens.
- Where did it connect? Management systems, package registries, and any destination that appeared after an update.
- Who started it? A user, an IDE, or an agent starting the server 162 times a week with nobody watching.
- What changed since yesterday? A new dependency set, a new outbound destination, or a new child process.
All five are answered on the endpoint, by watching processes, file access and network activity in context.
A practical checklist
- 1.Freeze high-privilege MCP servers the way Corelight did: pin to a commit, install from a hash-pinned lock file, launch directly, disable update checks.
- 2.Inventory MCP servers by behavior, not only by config, including servers started by IDEs and agents.
- 3.Alert on drift: a new binary or package hash under an approved server name, or a new outbound destination after an update.
- 4.Watch credential reads. A server or interpreter that reads several credential locations and then connects to a new host should be a high-priority signal.
- 5.Keep admin and SOC workstations stricter. Servers with reach into the firewall or SIEM deserve allowlisted egress and posture checks.
- 6.Rotate AI tool tokens on any device or repository touched by the wfx-loot campaign, and treat reachable cloud roles as exposed.
Where Kitecyber fits
Kitecyber runs on the endpoint and combines device posture, operating system and process activity, user activity, network activity and data classification. For MCP servers, that means seeing which servers and agents actually run on each device, what started them, which files and credentials they read, and where they connect, so drift shows up as a change in behavior rather than going unnoticed in a config file. Kitecyber can apply allow, warn or block policy when sensitive data or credentials head somewhere new. It complements dependency pinning and supply chain controls: they decide what should run, and Kitecyber shows what did.
Book a Kitecyber walkthrough at kitecyber.com to see the MCP servers and AI tools running across your endpoints today.

Ajay Gulati
Ajay Gulati is a passionate entrepreneur focused on bringing innovative products to market that solve real-world problems with high impact. He is highly skilled in building and leading effective software development teams, driving success through strong leadership and technical expertise. With deep knowledge across multiple domains, including virtualization, networking, storage, cloud environments, and on-premises systems, he excels in product development and troubleshooting. His experience spans global development environments, working across multiple geographies. As the co-founder of Kitecyber, he is dedicated to advancing AI-driven security solutions.