MCP Governance in 2026: How to See and Control the MCP Servers Your Employees Use
.jpg)
Direct answer: MCP governance is the practice of knowing which Model Context Protocol (MCP) servers your employees' AI tools connect to, deciding which ones are allowed, and controlling what data goes to them. For the servers you run, that means hardening them; for the far larger set you don't run, it means governing the connection from the managed laptop, because MCP traffic is ordinary outbound HTTPS that DNS filters can't read and cloud proxies only see when it's routed and decrypted through them.
Short answer: You cannot govern the Model Context Protocol (MCP) from the network edge, because MCP traffic is just outbound HTTPS to servers your tools never catalogued. DNS filters miss it, cloud proxies backhaul it, and neither can tell a sanctioned MCP server from a personal one. dope.security governs MCP on the device: the dope.security endpoint sees the connections MCP clients make to MCP servers, the MCP Usage report gives you visibility of which servers are in use, and the Fly Direct SWG web policy can allow, warn, or block their domains, with Dopamine DLP inspecting uploaded files or AI prompts in covered apps.
Six months ago, almost nobody in your company had heard of MCP. Now your developers are wiring Claude, Cursor, and a dozen internal agents into MCP servers that read files, hit APIs, and move data on their behalf. That is real productivity. It is also a brand new egress path that most security stacks cannot see, and it showed up without a single procurement ticket.
This is the same pattern we wrote about when we found 56 MCP domains hiding in plain sight across real customer traffic. MCP is the new shadow IT. The difference from the last shadow-IT wave is that these connections are designed to move data and take actions, not just render a web page. If you want the full picture of where this fits, our complete guide to AI visibility and governance is the hub. This post is about the MCP-shaped hole in it.
Last updated: November 2026
Key takeaways
- MCP governance has three parts: discovery (which servers are in use), policy (which are allowed), and data control (what is sent to them).
- Shadow MCP, the MCP servers employees connect without approval, is its own category in the OWASP MCP Top 10 (MCP09).
- Harden the MCP servers you run with the NSA and MCP-specification checklist. Govern the ones you don't run from the managed endpoint.
- DNS filters see hostnames only. Cloud proxies see MCP traffic only when it is routed through them and TLS-inspected.
- Endpoint tools see traffic, not intent: they can show that an MCP client reached a server, not which tool call an agent made.
What is MCP, and why did it become a security problem overnight?
The Model Context Protocol is an open standard that lets an AI model or agent connect to external tools and data sources through MCP servers. A server might expose your file system, a Jira instance, a database, a payments API, or a third-party SaaS. The model calls the server, the server does the work, and results flow back into the model's context.
That design is why MCP is useful and why it is risky. An MCP server is a live connection between a model and something that holds your data. It runs over ordinary HTTPS, often to a domain your security team has never reviewed, sometimes to a server a developer spun up on their own laptop. Nothing about that traffic looks different from any other encrypted request. That is exactly the problem.
What is MCP governance?
MCP governance is the set of controls that decides which MCP servers your AI tools may connect to, who may use them, and what data may flow through them, with monitoring to prove it. It is the MCP version of shadow IT governance: an inventory first, then policy, then data controls.
Employees add an MCP server with a single JSON entry in a client config file, with real credentials and no change ticket. That is why governance has to start with an inventory rather than a policy document.
What is shadow MCP?
Shadow MCP is any MCP server an employee connects to an AI tool without IT or security approval. The OWASP MCP Top 10 lists Shadow MCP Servers (MCP09) as a category of its own: unapproved MCP deployments operating outside formal security governance. It is the shadow IT problem, one layer down. (More in What Is Shadow MCP?)
Shadow MCP shows up in a few common shapes:
- Desktop agents with connectors: Claude Desktop or ChatGPT desktop with MCP connectors to Drive, Slack, or GitHub.
- Coding agents: Claude Code, Gemini CLI, and IDE agents that run shell commands and launch MCP servers.
- Local MCP servers: file system, GitHub, or Postgres servers running on the laptop and reaching remote APIs.
- Remote MCP connectors: hosted connectors run by a SaaS vendor, an open-source maintainer, or a stranger on GitHub.
The difference from chat-style shadow AI is blast radius. A pasted prompt leaks what the user typed. An agent with a repo token or Files.ReadWrite.All can read and change everything the user can. For a layer-by-layer look at how to find them, see shadow AI agents: tools to detect agents and MCP.
What are the main MCP security risks?
Drawn from the OWASP MCP Top 10, the MCP specification's security best practices, and the NSA's May 2026 guidance:
- Token mismanagement and secret exposure (MCP01): long-lived tokens and API keys stored in client config files, model memory, or logs. GitGuardian's State of Secrets Sprawl 2026 found 24,008 unique secrets exposed in MCP configuration files.
- Privilege escalation via scope creep (MCP02): a server granted broad scopes such as
files:*oradmin:*up front, so one stolen token reaches everything. - Tool poisoning (MCP03): a compromised or malicious server feeds misleading tool descriptions or outputs that steer the model. A Cloud Security Alliance research note also describes IDEs that auto-executed project-defined MCP servers with developer-level privileges.
- Command injection (MCP05): the agent builds and runs commands from untrusted input returned by a tool.
- Prompt injection via context (MCP06): data a tool returns contains instructions the model follows.
- Insufficient authentication and authorization (MCP07): unauthenticated remote servers, token passthrough, and the confused-deputy attack on MCP proxy servers that the specification describes in detail.
- Lack of audit and telemetry (MCP08): no record of which tool was invoked, by whom, with what.
- Shadow MCP servers (MCP09) and context over-sharing (MCP10): servers nobody approved, and persistent context leaking one user's data to another.
The NSA's Artificial Intelligence Security Center put it bluntly in its May 20, 2026 Cybersecurity Information Sheet: established cyber defense strategies do not adequately address the new risks MCP introduces. Researchers filed more than 30 CVEs against MCP servers, clients, and infrastructure in January and February 2026 alone, according to Cycode's analysis of the OWASP list. The protocol is young and the implementations are younger.
How do you secure the MCP servers you run?
If your company hosts MCP servers for employees or customers, this is the baseline the NSA and the specification set out. Our 2026 checklist for securing the MCP servers employees connect to AI tools walks through each item with its NSA and OWASP sources.
- Authenticate every remote server. Use OAuth 2.1 with PKCE. Never expose an unauthenticated MCP endpoint on a network.
- Validate token audience and never pass tokens through. A server must reject any token not explicitly issued to it, or it becomes a confused deputy for downstream APIs.
- Minimize scopes. Start with a low-risk discovery scope and step up only when a privileged tool is first called. Don't publish wildcard scopes.
- Implement per-client consent on proxy servers. Show your own consent screen per registered client, validate redirect URIs with exact matching, and use single-use OAuth state values.
- Sandbox tool execution. Restrict file system and network access, type-check every parameter, and treat tool output as untrusted input.
- Sign and log everything. Sign and verify messages where supported, and log every tool and model invocation with the identity behind it.
- Keep internal servers off public networks and rate-limit anything reachable from the internet per identity.
- Harden local servers. Prefer the stdio transport for servers on the laptop, require a token on any local HTTP transport, and show users the exact command a one-click install will run.
- Audit the code and the supply chain. Choose supported MCP projects, review server code before deployment, and pin versions.
Why do DNS filters and cloud proxies miss MCP traffic?
Most teams assume their existing web security already covers this. It usually does not. A DNS-layer tool like Cisco Umbrella resolves a domain and moves on. It never sees the URL path, the payload, or the tool call inside the session, so it cannot tell you what an MCP server did, only that a name was looked up. Cisco's own documentation is clear that tenant-aware control needs a proxy with SSL decryption, not DNS. When roughly 95% of traffic is encrypted, DNS-only visibility into MCP is close to zero.
Cloud proxies see more, but they add their own tax. Every MCP call from a developer's machine has to detour to the vendor's nearest point of presence and back before it reaches the server. For latency-sensitive agent loops that fire dozens of calls, that round trip is felt. Worse, many cloud proxies bypass inspection on cert-pinned developer tools by default, which is precisely where MCP clients live. The tool you most need to inspect is the one on the blind list.
Cisco's own MCP and agent gateway controls, announced in March 2026, live in Secure Access rather than Umbrella, and Umbrella reaches end of sale on January 31, 2027.
Cloud proxies can inspect MCP traffic if it is routed through them and TLS-inspected, and most now sell an AI or agent product to do it. Zscaler announced AI Broker at Zenith Live in June 2026 and Netskope announced Skylight Agentic Broker on September 15, 2026, both without disclosed GA dates or pricing at announcement. Cloudflare's shadow-MCP detection (August 14, 2026) only works on traffic Gateway already TLS-inspects. In each case every MCP call detours to a point of presence and back, and cert-pinned developer tools on a bypass list are not inspected, which is exactly where MCP clients live.
What does governing MCP actually require?
Governing MCP is not one control. It is three, and they have to work together on the same egress point. First, discovery: you need to see every MCP server your people connect to, sanctioned or not. Second, policy: you need to allow the servers you approved and block the rest, by name, not by a blunt category. Third, data control: you need to inspect what leaves in the call itself, because the risk is not the connection, it is the customer record or source file that rides along inside it.
Here is how the common approaches stack up against those three requirements.
| MCP governance need | dope.security | DNS filter (e.g. Umbrella base) | Cloud proxy (e.g. Zscaler, Netskope) |
|---|---|---|---|
| Discover every MCP server in use | Yes: Shadow IT discovery on the device | Partial: domain lookups only | Partial: if not on a bypass list |
| Allow sanctioned, block the rest by name | Yes: SWG policy + Cloud Application Control | No: cannot read URL or tenant | Add-on, higher tier |
| Inspect data sent to AI apps | Yes: Dopamine DLP on uploaded files or AI prompts in covered apps (not MCP tool calls) | No | Separate DLP SKU |
| No detour on every agent call | Yes: inspection on-device, fly direct | N/A | No: backhaul to a PoP |
Which MCP servers are your employees using? The MCP Usage report
On October 29, 2026, dope.security released the MCP Usage report. It lets you track which MCP servers are being used across your org, by whom, and how much data is flowing through. MCP Usage is a visibility report, and you find it in dope.console.
That gives you an inventory built from real traffic instead of a survey. Use it as the first step of the plan below: sort the servers you find into approved, tolerated, and blocked, then come back to it regularly while adoption is moving fast.
One honest note on scope: endpoint tools see traffic, not intent. They can tell you an MCP server is talking to a remote host, but not which tool call the agent made. dope.security doesn't parse individual MCP tool calls or provide runtime agent guardrails, so pair it with an MCP allowlist and, for agents you build, an MCP or AI gateway. The report is for visibility; enforcement happens in your web policy. We compare what the endpoint, OAuth discovery, identity tools, and AI gateways each see in our guide to tools that detect shadow AI agents.
How does dope.security govern MCP on the device?
Because dope.security runs an agent on the endpoint and inspects SSL there, the connections MCP clients make to MCP servers are visible at the point they leave the machine. There is no data center to route through and no domain that slips by because it was never in a category list. The same three-layer model we use for AI governance applies directly to MCP.
Layer one is discovery. Shadow AI and shadow IT discovery surfaces the MCP servers your people are actually reaching, including the ones nobody told you about. You get names, not guesses. Layer two is SWG policy plus Cloud Application Control: allow the MCP servers you sanctioned, block the personal or unknown ones, and do it per user or group with policy that pushes in seconds. Layer three is Dopamine DLP, which inspects uploaded files or AI prompts in covered apps in Monitor or Block mode; it does not inspect MCP tool calls or tool arguments. The classification runs through a zero-retention API, so inspecting the data does not create a second copy of it somewhere else.
The dope.security endpoint sees outbound connections, including from the MCP client inside Claude Desktop, Cursor, VS Code, or a terminal. On the data side, Dopamine DLP extracts text from the prompt or file on the device and sends only that text to dope.security's cloud, where a zero-data-retention LLM API classifies it. Covered AI apps include ChatGPT (prompts and file uploads), Anthropic/Claude (desktop app, CLI, VS Code, Cursor, Cowork, Microsoft Add-ins), Gemini, Microsoft Copilot (browser), Perplexity AI (browser), Abacus AI (browser), and Google NotebookLM.
For developer-heavy fleets, see shadow AI tools for Claude Code, Cursor, and developer AI, and for why on-device inspection matters here, on-device architecture for shadow AI and shadow MCP.
Why is the payload the point?
It is tempting to treat MCP governance as an allow-list exercise. Approve a few servers, block the rest, done. That misses the real exposure. An approved MCP server can still be handed data it should never receive, because the model decided that reading a spreadsheet of customer records was the fastest way to answer a question. The connection was fine. The payload was not.
This is why inspecting the traffic beats cataloguing the domains. MCP Usage gives you visibility of the servers, not the tool calls, so pair it with Dopamine DLP on the AI apps your people use: if a prompt or uploaded file carries PII, PCI, PHI, or source code, Dopamine DLP can monitor or block it, depending on the policy you set. For coaching, Warn mode is a Cloud Application Control feature at the access level for Gemini and Copilot: users see a warning page and can choose to proceed; admins can customize the page title and message. You govern the data, not just the destination. Our take on sanctioned versus unsanctioned SaaS makes the same argument: the line between approved and personal is where most tools quietly give up.
What is a practical starting point for MCP governance?
You do not need a committee to begin. Turn on discovery and look at which MCP servers are already in use. You will almost certainly find some you did not know about. Sort them into sanctioned and everything else. Put the everything-else group behind a block or a warning while you review it. Then turn on data inspection for the sanctioned servers, because those are the ones with standing access to real systems. That sequence, see it, sort it, inspect it, takes days on dope.security, not a quarter, because there is no proxy to stand up and no PoP to route through.
MCP is not going away, and blanket-banning it just pushes your best engineers onto personal machines where you have no visibility at all. The winning move is to make the sanctioned path the easy one and watch the rest.
What does MCP monitoring look like?
MCP monitoring is the ongoing half of governance: watching which servers are in use, who uses them, and how usage changes, then feeding that back into policy. A workable monitoring loop has four parts:
- A regular inventory of which MCP servers are in use and by whom, from the MCP Usage report.
- An audit trail of web traffic and policy decisions sent to your SIEM through dope.security's SIEM integration, so the client side of OWASP MCP08 has a record.
- Server-side logs for any MCP servers you host, with the identity behind every tool invocation.
- A review trigger: any new MCP server in the inventory goes through the approved, tolerated, or blocked decision before it spreads.
What does a 30-day MCP governance plan look like?
- Days 1 to 7, discover. Deploy dope.security to a pilot group via Intune or Jamf and read the MCP Usage, Shadow IT, and AI Usage Analytics views. List every MCP server and who uses it. Our list of 56 common MCP domains is a starting map.
- Days 8 to 14, classify. Sort servers into approved (vendor-run, authenticated, scoped), tolerated (open source, reviewed), and blocked (unknown, unauthenticated, remote with broad scopes). Apply the server-side MCP checklist to anything your own teams host.
- Days 15 to 21, enforce. Allow approved servers, warn on tolerated ones, block the rest. Warn and block here are web policy on the server's domain. Separately, run Dopamine DLP in Monitor for covered AI apps, then switch to Block once you've reviewed detections and set exceptions.
- Days 22 to 30, prove it. Send dope.security logs to your SIEM so connections and policy verdicts have an audit trail, and write the MCP section of your AI acceptable use policy from what you actually found.
For the broader framework, see agentic AI security and the OWASP MCP Top 10 explained for IT teams.
When should you choose dope.security for MCP governance?
Choose dope.security when your employees use AI clients on Mac and Windows laptops you manage, when you can't inventory the MCP servers they have connected, and when the proxy vendor's answer is an announced-but-unpriced agent broker that still routes every request through a data center. Greylock Partners moved off DNS-only filtering for the same reason: the risk had moved inside HTTPS, and DNS couldn't follow it.
Choose a server-side MCP gateway alongside it when your company builds and hosts MCP servers at scale and needs centralized OAuth, scope management, and tool signing. The two layers are complementary: one hardens what you run, the other governs what your people connect to. dope.security covers managed Mac and Windows endpoints, not personal phones or unmanaged laptops.
Frequently Asked Questions
What is MCP server security?
MCP server security is the practice of discovering, controlling, and inspecting the Model Context Protocol connections that AI models and agents use to reach external tools and data. Because MCP runs over ordinary HTTPS to servers that are often uncatalogued, effective MCP security means seeing every server in use, allowing only sanctioned ones, and controlling the data employees send to AI apps. dope.security addresses all three from the endpoint: MCP Usage for visibility of the remote MCP servers your employees' AI clients connect to, Fly Direct SWG web policy to allow, warn, or block domains, and Dopamine DLP in Monitor or Block mode for uploaded files or AI prompts in covered apps.
Can a DNS filter or firewall control MCP traffic?
Not meaningfully. A DNS filter only resolves domain names, so it cannot read the URL, the tenant, or the payload of an MCP call, and most MCP traffic is encrypted. A traditional firewall has the same blind spot. Controlling MCP requires on-device SSL inspection where the connection starts. The dope.security endpoint sees the connections MCP clients make to MCP servers, and web policy can allow, warn, or block their domains.
How is governing MCP different from blocking a website?
Blocking a website is a yes-or-no decision about a destination. Governing MCP is about the data and actions a live connection carries, because an approved MCP server can still be handed customer records or source code by a model trying to be helpful. That is why dope.security pairs MCP Usage, which gives visibility of servers rather than tool calls, with Dopamine DLP on uploaded files or AI prompts in covered apps, rather than only allowing or denying the domain.
Does inspecting MCP traffic slow down AI agents?
With a cloud proxy it can, because every call detours to a point of presence and back. dope.security inspects on the device and flies direct, so there is no network detour added to agent loops that may fire dozens of MCP calls in sequence. Inspection happens where the traffic already is.
What data can leak through an MCP server?
Anything the model can reach: PII, payment data, health records, intellectual property, and source code, all of which can be passed into an MCP call as context or as a tool argument. dope.security does not inspect MCP tool calls or tool arguments. For the AI apps your people use, Dopamine DLP extracts text from the prompt or file on the device and sends only that text to dope.security's cloud, where a zero-data-retention LLM API classifies it, and it can monitor or block based on your policy.
What is MCP governance?
MCP governance is the set of controls that decides which Model Context Protocol servers your AI tools may connect to, who may use them, and what data may flow through them, backed by monitoring. In practice it means discovery, an allowlist, data controls, and an audit trail.
What is shadow MCP?
Shadow MCP is any MCP server an employee connects to an AI tool without IT or security approval. OWASP lists Shadow MCP Servers (MCP09) in its MCP Top 10. Because a single config entry can connect an agent to production systems with real credentials, shadow MCP belongs in your shadow AI inventory and allowlist.
What are the biggest MCP security risks?
The OWASP MCP Top 10 lists token and secret exposure, scope creep, tool poisoning, command injection, prompt injection via context, weak authentication, missing audit trails, shadow MCP servers, and context over-sharing. GitGuardian found 24,008 unique secrets exposed in MCP configuration files in its State of Secrets Sprawl 2026 report.
How do I find out which MCP servers my employees are using?
Use observed traffic, not a survey. dope.security's MCP Usage report, released October 29, 2026, lets you track which MCP servers are being used across your org, by whom, and how much data is flowing through. It is a visibility report in dope.console; allowing, warning, or blocking is done separately in web policy.
Does dope.security replace an MCP gateway for servers we host?
No. dope.security governs the client side: which MCP server domains managed laptops may reach, and what data goes into covered AI apps. If you host MCP servers, you still need OAuth 2.1, scope minimization, and tool sandboxing on the server, built in or through an MCP gateway.
See it on your own traffic. Start a free trial or book a 20-minute demo and watch dope.security surface the MCP servers already running in your environment.


.jpeg)
.jpeg)
.jpeg)

