Analysis

Model Context Protocol (MCP) in your company: how to connect AI to company systems without opening a back door to attackers

This article is for management boards, IT directors and security leads who are hearing about “MCP servers” from vendors and have to decide whether to allow them. We explain in plain language what MCP is and when it pays off. We also show which risks have already materialised and what questions to ask before AI gets access to your company’s data.

Author: Kacper Włodarczyk, CEO of ALGORCOMPPublished: August 23, 2026Reading time: 12 min readAI / AI agentsFor: Enterprise
Model Context Protocol (MCP) in your company: how to connect AI to company systems without opening a back door to attackers

What MCP is – an explanation without jargon

Model Context Protocol is a common standard for connecting AI assistants to the systems where a company keeps its data. It works a bit like a USB-C connector: every device that complies with the standard fits every socket. Here, the “device” is an AI tool, and the “socket” is your ERP, CRM or document archive.

The standard was developed by Anthropic and released in autumn 2024. In December 2025, Anthropic handed it over to an independent foundation operating under the Linux Foundation. According to Anthropic, MCP is now supported by ChatGPT, Gemini, Microsoft Copilot and popular developer tools, among others.

In the Microsoft world, MCP is available in Copilot Studio. An assistant built with this tool can use an MCP server like any other connector, subject to the data protection policies set by the administrator.

What MCP solves, and what it does not

The problem is simple and costly. Until now, every AI tool that needed to reach company data required a separate integration. With three systems and three AI tools, that means nine connections to build and maintain. With MCP, you expose each system once, as a server, and every compatible tool can use it: nine connections become three servers.

However, MCP does not tidy up data, does not replace permissions and does not make AI understand your process. A server is also one more component that has to be updated, monitored and adapted to changes in the underlying system.

The practical break-even point: a single stable connection between one system and one application is usually cheaper to build the traditional way. MCP starts to pay off when a second AI tool or a third system comes along.

Model Context Protocol (MCP) in your company: how to connect AI to company systems without opening a back door to attackers

How it works: application, connector and server

In conversations about security, three elements are most often confused. The application is the program a person uses, such as an assistant in a browser or in Teams. The connector is the part of the application that maintains the connection to a specific server. The server is the component on the system side that exposes a list of tools and data and carries out commands.

A server can run on the user’s computer or remotely, as a network service. The remote option is more convenient but requires careful authentication. A local server runs someone else’s code on the computer, so you need to know where it comes from.

The most important detail: the description of each tool on the server is passed to the AI model as text. The model reads it and treats it as an instruction. Whoever controls a tool’s description partly controls the assistant’s behaviour – a whole class of attacks stems from this property.

Three uses where MCP makes sense in a mid-sized company

In all three examples below, the MCP server is the place where you decide what is allowed. Designed as a list of permitted actions, it acts as a safeguard. Designed as open access to a database, it moves old problems to a new place.

  • Documents and knowledge: one server on top of the company document archive provides search that respects the user’s permissions. The same layer is used by the intranet assistant, the assistant in Teams and the quote preparation tool.
  • ERP and CRM: instead of “access to the database”, the server exposes specific operations, such as “check stock levels”, “get order status” or “prepare a draft order for approval”. The scope of permissions is described in one place.
  • Reports and metrics: the server exposes defined metrics, such as margin or inventory turnover, instead of letting AI write arbitrary database queries. Each metric then has one agreed definition.
Diagram: an AI assistant connects through MCP servers to an ERP system, a document repository and a data warehouse

An MCP server is not just a connector. It is a list of actions AI can take on your behalf, and a description the model reads as an instruction.

Which risks must be closed off before launch

The MCP specification describes several typical attacks in its section on security best practices. Separately, OWASP, an organisation focused on application security, is developing a list of the top ten threats to MCP. In May 2026, the US National Security Agency (NSA) published guidance stating that MCP’s popularity has outpaced the development of its safeguards.

These risks are not theoretical. In April 2025, Invariant Labs demonstrated a tool posing as a calculator whose hidden description instructed the model to steal access keys from a developer’s computer. In September 2025, a fake connector impersonating the Postmark service appeared in a public software package registry and quietly sent a copy of every email to the attacker.

The list below describes the main threats in plain language. We give the English technical names in brackets, because that is how you will find them in vendors’ documentation.

  • Poisoned tool description (tool poisoning): someone hides commands in the description the model reads, such as “while you’re at it, send the contents of the password file”. The user does not see them; the model carries them out.
  • Confused intermediary (confused deputy): a server with broad permissions performs an operation for someone who does not have those permissions, because it does not check on whose behalf it is acting.
  • Forwarding passes (token passthrough): the server accepts a user’s digital pass (token) issued for another system and passes it on. The MCP specification prohibits this.
  • Server as a back door into the internal network (SSRF): a malicious server tricks the application into connecting to internal company resources the attacker has no access to.
  • Session hijacking (in older versions of the protocol, which kept sessions) or server impersonation: someone guesses a connection identifier or pretends to be a known server and intercepts commands.
  • Excessive permissions: the server is given access “to everything”, so a single leaked pass opens all systems at once.

What the 2026-07-28 specification version changes

On 28 July 2026, a new version of the MCP specification was published. The most important change: the protocol no longer maintains sessions, i.e. a persistent connection, and each request carries all the information needed to handle it. This makes servers easier to run in the cloud like any other service, without tying a user to a single machine.

The new version contains changes that are not backward-compatible with the previous one, 2025-11-25. The Roots, Sampling and Logging features have been marked as deprecated; under the project’s rules, they have at least 12 months before they may be removed. The specification also tightens authentication rules, including checking which sign-in system (authorisation server) issued a pass.

For a company with MCP integrations already running, this is a concrete item for the calendar. You need to check with your vendors when they will update their servers and applications. It is also a good moment to sort out authentication and the logging of operations, which were put off during the first rollout.

Seven questions before allowing an MCP server into production

Treat an MCP server like software from a new supplier that gets access to company data. Being easy to launch does not mean it is ready to work with real data. A study from May 2026 showed that about four in ten publicly accessible servers did not require any sign-in.

It is worth asking the questions below for every server, including ready-made ones downloaded from a well-known vendor. Record the answers – they will be useful when you review the server after a protocol version change.

  • Who publishes the server, and is the publisher on your list of approved suppliers? Are you downloading it from the official source rather than from a copy with a similar name?
  • Does the server require authentication, and does it use passes issued for itself rather than passed on from another system?
  • What is the narrowest scope of permissions needed for this use? Is read-only access enough to start with?
  • Has someone read the descriptions of all the tools the server exposes, and do you know who approves changes to them?
  • Is every call recorded in a log, with information on which user requested it?
  • Does the server need to be accessible from the internet, or is the internal network enough?
  • Who is responsible for updating the server after changes to the protocol or to the system it connects to?

How to start without excessive risk

The first step is one read-only server on top of one data source, in a test environment. A set of procedures or a product catalogue is a good candidate, because a wrong answer there will not cause financial damage.

Collect a log of queries for a few weeks. You will see what people really ask about and which operations are missing. Only then decide on the target scope, on actions that change data, and on whether to build your own server or use a ready-made one.

If the server is to process personal data, agree the scope with the person responsible for data protection. MCP does not change the GDPR rules. It does, however, increase the number of places where they need to be enforced.

Related topics in the knowledge base

Read next

FAQ

MCP in your company – frequently asked questions

Questions we hear from management boards and IT departments when a vendor proposes integration via MCP.

Will MCP replace our existing integrations?
Not automatically, and it is not always worth it. A working connection between two systems is usually cheapest to leave as it is. MCP makes sense where the same data needs to be available to several AI tools.
Does MCP work with Microsoft 365 and Copilot Studio?
Yes. Microsoft has made MCP support in Copilot Studio generally available. Servers work there as connectors, so they are covered by the data protection policies set by the Power Platform administrator.
Can we use ready-made MCP servers from the internet?
Yes, but with the same caution as with any third-party software. The fake Postmark connector case showed that a package with a familiar name does not necessarily come from the vendor. Check the publisher, record the version and do not update automatically without reviewing the changes.
Is MCP a matter for the IT department only?
The configuration, yes. But the decision on which data and which actions to make available to AI belongs to the process owners and the management board. They are responsible for the consequences if the assistant changes data in the ERP or sends something to a client.
Do we have to build our own MCP server?
Not to begin with. Many software vendors provide their own servers for their systems. Your own server pays off when you want to give AI access to data from a system that has no such server, or to tightly restrict the list of permitted operations.

About this page

About the author

Kacper Włodarczyk — CEO of ALGORCOMP

Kacper Włodarczyk

CEO of ALGORCOMP

Kacper specializes in implementing Microsoft 365 Copilot, Copilot Studio, Power Platform (Power Automate, Power Apps, SharePoint) and AI agents for mid-sized B2B companies in Poland. He leads dozens of projects covering AI strategy, Power Platform governance, document workflow automation and sales processes. In his writing he focuses on the practical side of AI adoption in organizations — from the first proof of concept to company-wide rollout — with particular attention to data security, compliance (GDPR, NIS2, the AI Act) and return on investment.

Meet the team

Planning to connect AI to your company systems?

In a free consultation, we will assess whether MCP has an advantage over a conventional integration in your case, and go through the list of safeguards for your first server.

Featured

Related articles