Blog · Opinion · 18 November 2026

Closing the doors on MCP and calling it security

Microsoft now sells enterprises a firewall for MCP servers, and when the outside servers are switched off the agents are left with Microsoft's own tools and Microsoft's index of the web.

10,000active MCP servers when Anthropic gave MCP to the Linux Foundation in December 2025
163MCP servers in our directory on 15 November 2026, each graded from public evidence
6things Microsoft's MCP Firewall controls, from servers and tools to protocols and transports
11 August 2025the day Microsoft switched off the Bing Search APIs for everyone outside its agent platform
MCP, open and closed. Sources [1], [2] and [3], and our listings.

Microsoft used Ignite, which opened on 17 November 2026, to sell enterprises a way to shut the door on MCP servers they didn't choose. Entra gets an MCP Firewall, in preview, that controls access to MCP servers, tools, resources, prompts, protocols and transports [1]. Agent 365 will let admins watch local and shadow MCP servers, the ones nobody approved, and pick out tools to investigate, and Microsoft promises finer controls than blocking them later, so blocking is where it starts [1]. Intune will find the agents running on a company's laptops, mark any device with an unapproved agent as noncompliant and use Conditional Access to cut it off from company data [1]. Entra will give an Agent ID to every local agent it discovers, starting with GitHub Copilot CLI and Claude Code [1].

I think this is a model trainer and a web indexer using the word security to keep its moat, and I think Google is quite happy to watch Microsoft make the argument, because it sells the same allowlists. The security problem is real. The answer they're selling is that the vendor who owns the index decides what the enterprise agent is allowed to read.

The risk is real

An MCP server hands a model a list of tools, and the descriptions in that list are text the model reads and often follows. A server can hide an instruction in a tool's description, ask for more than it needs or change its tools after you've approved it, and a local agent runs with the user's own access to files and accounts. Microsoft says the same in its own words, that malicious instructions hidden in the content an agent processes can push it into actions nobody asked for. Defender now scans local agent calls for them on Windows, in Claude, ChatGPT, GitHub Copilot, Ollama and others [1]. We built check_server into our MCP server for the same reason, to read a public server's tool list from outside and say what's aimed at the model, which safety hints are missing and what it costs in tokens, without calling a single tool.

So I don't think a company should let every engineer plug any server into an agent that can read the finance share. I think the question is who decides what goes on the list, and on what evidence.

Who sells the lock

Microsoft sells the firewall, and the governance that makes it useful needs a licence. "An Agent 365 license is still required for customers to get enterprise-grade management, security and governance" [1]. The same week it made Microsoft MCP Server for Enterprise generally available, which lets agents query a company's Entra directory [1]. It also launched Microsoft IQ, which connects agents to work data, business data and the web through Work IQ, Fabric IQ, Foundry IQ and Web IQ [1]. Those are the tools that stay switched on when the outside ones are blocked, and they come from the same vendor that runs the block.

Web IQ is the part that matters most. It's Microsoft's web grounding for agents, and Microsoft describes it as "built on twenty years of Bing search infrastructure" [5], starting from the Bing global index [4]. At Ignite it became generally available to all Azure customers [1]. Microsoft retired the Bing Search APIs on 11 August 2025, decommissioned every instance and told customers to move to Grounding with Bing Search inside Azure AI Agents [3]. Grounding with Bing Search costs $14 per 1,000 transactions [9], and developers never see its raw results, only the model's answer with links to the pages it used [10]. It's in our directory as Grounding with Bing Search.

Put those together and an enterprise that follows Microsoft's advice ends up with agents that reach the web through Bing's index, inside Azure, priced by Microsoft. Every independent search, scraping or data server that might have given a second view is a shadow MCP server waiting for investigation. Microsoft says it's a security feature, and I think it's also a distribution deal for Bing that the customer pays for.

Google sells the same thing

Google's version is already shipping. In Gemini Enterprise the setting for MCP servers is off by default, and an admin who turns it on writes an allowlist of local and remote servers [6]. Gemini CLI's enterprise controls work the same way, and once the allowlist has a server in it "all locally configured servers not present in the allowlist are ignored" [7]. Google Cloud's own MCP servers, for BigQuery and the rest, come with IAM allow and deny policies and Model Armor scanning [8]. I don't know what Google's product team thought watching Microsoft's keynote, but every enterprise that learns to call an unlisted MCP server a threat is an easier sale for Google's own search grounding as well.

Both companies backed the Agentic AI Foundation when Anthropic gave MCP to the Linux Foundation, so that MCP would stay open and vendor-neutral [2]. That was eleven months ago. Now they sell the firewall that decides which parts of the open protocol your agents may use.

Open is how it got safe

Open data, open code and open weights are what make society better, and they're also what these companies run on. Kubernetes, PyTorch and Node.js sit under the same Linux Foundation stewardship as MCP [2], and the agents being fenced in speak a protocol their makers gave away. Open code got safe because anyone could read it, and that's the model I'd want for agent tools too, with tool lists that are public, readable from outside and graded by someone who doesn't sell the alternative.

That's what we do with the 163 MCP servers in our directory. Each one is graded from public evidence against a published checklist, and anyone can read its tool list with check_server, so the evidence is there for any admin to use or to argue with. An allowlist built from that is a security control, and one built from the vendor's own catalogue mostly protects the vendor.

If I were running IT at a company today, I'd keep the allowlist and change what goes on it. Read every server's tool descriptions before it's approved, pin the version you approved, prefer servers with public code and run open weights on your own hardware for the work that can't leave the building. Above all, don't let the company that sells you search be the one deciding which search your agents can use.

Where I am

I think Microsoft is right that agents need identities, logs and limits, and wrong that the limits should come from the vendor with the most to gain from them. Disclosure, we grade MCP servers for a living, so a world where enterprises read them openly is good for us. Microsoft's and Google's tools are in our directory under the same checklist as everyone else's, and our research agents and review panel run on Anthropic's Claude, which is one of the first two local agents Entra will give an ID.

I'd bet Microsoft's own MCP server is on most of those allowlists on day one. Whether anyone else's gets on is up to the admin, and I'd rather that admin had read the tool list than the brochure.

References

Sources read on 18 November 2026.

  1. [1]Microsoft, Microsoft Ignite 2026 Book of News, 17 to 20 November 2026. Source for MCP Firewall and local agent controls in Entra Global Secure Access, Agent 365's governance for MCP servers and for local and shadow MCP servers, the Agent 365 licence for enterprise-grade governance, Intune's local agent governance with Conditional Access, Agent ID for discovered local agents starting with GitHub Copilot CLI and Claude Code, Defender's runtime protection for local agents, Microsoft MCP Server for Enterprise, Microsoft IQ and Web IQ's general availability to Azure customers. news.microsoft.com
  2. [2]Model Context Protocol blog, MCP joins the Agentic AI Foundation, 9 December 2025. Source for 10,000 active servers, the donation to the Linux Foundation's Agentic AI Foundation, its support from Google and Microsoft, vendor neutrality, and Kubernetes, PyTorch and Node.js under the same stewardship. blog.modelcontextprotocol.io
  3. [3]Microsoft Lifecycle, Bing Search APIs retiring on August 11, 2025, 16 May 2025. Source for the retirement and decommissioning of the Bing Search APIs and the move to Grounding with Bing Search in Azure AI Agents. learn.microsoft.com
  4. [4]Microsoft Bing, Web IQ Blog, Announcing Microsoft Web IQ, June 2026. Source for Web IQ starting from the Bing global index. blogs.bing.com
  5. [5]Microsoft, Web IQ, frequently asked questions. Source for Web IQ being built on twenty years of Bing search infrastructure. webiq.microsoft.ai
  6. [6]Google Cloud, Gemini Enterprise, Configure AI developer tools settings. Source for the MCP servers setting being disabled by default and the admin allowlist of local and remote servers. docs.cloud.google.com
  7. [7]Gemini CLI, Enterprise Admin Controls. Source for the MCP server allowlist and locally configured servers outside it being ignored. geminicli.com
  8. [8]Google Cloud, Google Cloud MCP servers overview. Source for IAM allow and deny policies and Model Armor on Google Cloud's MCP servers. docs.cloud.google.com
  9. [9]Microsoft Bing, Grounding with Bing. Source for the price of $14 per 1,000 transactions. microsoft.com
  10. [10]Microsoft Learn, Use Grounding with Bing Search tools with the agents API. Source for developers and end users having no access to the raw content Grounding with Bing Search returns, and the model's answer carrying citations. learn.microsoft.com

Your experience matters

Used a tool in the Terminal?

Tell people and their agents what worked and what didn't. Write a review of a tool you've used and confirm it by email or wallet.