How MCP is creating a new communication layer - and a new security blind spot
MCP is becoming a common communication layer for AI agents, giving them a standard way to discover and use tools, access data and take actions across applications and services. But every new way of connecting also creates a new way in. As organizations expose remote MCP servers to make their capabilities available to agents, they are opening a new network-facing surface that needs to be protected. And while much of the MCP security conversation has focused on protecting the agent, there is another side of the connection that deserves attention: who is protecting the MCP server?
When applications started talking to applications
For most of the last two decades, applications talked to applications through APIs.
A mobile application called a payment service. A CRM retrieved customer data. A partner checked inventory or submitted an order. These interactions were built around defined interfaces and workflows, with the calling application generally doing what its developers had designed it to do.
Security evolved around the same model. Web application protection focused on browser traffic and application requests. API security followed as APIs became the standard way for applications and services to communicate, bringing new questions around endpoints, identities, authorization and data.
The pattern was familiar: a new way to connect applications created a new way to reach them, and security evolved alongside it.
Then the caller changed.
When the caller became an agent
AI agents introduced a different kind of application consumer.
Traditional integrations usually follow workflows defined by their developers. An agent can start with a goal and determine the steps needed to achieve it. It can discover available capabilities, choose which tool to use and use the result of one action to decide what to do next.
The system on the other side of the connection is no longer necessarily talking to another application following a fixed sequence. It may be talking to an agent that is making decisions at runtime.
This is what makes agents powerful, but it also creates a new requirement. Agents need a consistent way to discover and use capabilities across the applications and services they interact with.
MCP becomes the connection layer
This is where MCP fits into the picture.
The Model Context Protocol provides a standard way for AI applications to discover and use tools and resources exposed by an MCP server. Instead of building a separate integration between every agent and every application, an MCP server can expose capabilities through a common protocol.
The basic model is straightforward:
AI Agent → MCP Client → MCP Server → Tools, Data and Actions
The exposed tool definitions form a machine-readable catalog, often called a tool manifest. Tool names, descriptions and input schemas give an agent a map of what it can call and how to call it. A malicious agent can use that same map to select operations, probe their inputs and decide what to try next.
The MCP server becomes the interface between the agent and the systems behind it. It can expose existing capabilities without requiring the agent to understand how the underlying applications work.
As agents become more capable and more connected, that common interface becomes increasingly valuable.
The door opens
Once capabilities can be exposed through a common protocol, the number of things an agent can interact with can grow quickly.
A customer-facing application can expose tools for retrieving information or completing a transaction. A SaaS application can allow an agent to work with customer data. An internal application can make business functions available to an employee's AI assistant.
The use cases are different, but the model is the same: the agent needs a way in, and the MCP server provides it.
That is the opportunity. It is also where the security question begins.
For a remotely exposed MCP server, that interface is a network-facing entry point to tools, resources and actions.
The door has opened.
Everyone is watching the agent
The first security concerns around MCP naturally focused on the agent.
What happens if an MCP server provides malicious instructions? What if a tool description is designed to influence the model? What if an agent is manipulated into taking an action that its user never intended?
These are important questions, and they have become a major part of the MCP security conversation.
But they look at the connection from one direction:
MCP Server → Agent
There is another direction:
Agent → MCP Server
Here, the MCP server is not the source of the threat. It is the component being exposed.
And that distinction matters.
Who is protecting the MCP server?
The MCP server may expose access to data, tools and actions that were previously reachable only through other application interfaces. It is now part of the path between a caller and the systems behind it.
That makes it a new security boundary.
The caller may be an AI agent, a client acting on behalf of an agent, or another system using the protocol. Its behaviour may not be known in advance.
This does not mean the existing security model becomes irrelevant. It means the security decision needs more context. Seeing a request reach an MCP endpoint is not enough. Security needs to understand which Actor is making the request, which tool is being invoked, what the operation can do, and what workload it creates behind the server.
That is the new challenge MCP introduces: the endpoint may look the same while the actions behind it can be completely different.
A new security conversation
MCP is changing how applications expose their capabilities, just as APIs changed how applications communicated with each other.
But the same connection that makes those interactions possible also creates a new surface that can be reached and abused.
The conversation has largely focused on what a malicious MCP server can do to an agent.
We also need to look in the other direction.
What happens when the MCP server itself becomes the target?
In Part II, we take a closer look at the MCP server as an attack surface and examine the different ways it can be targeted.