The Hidden Concerns Behind the MCP Boom

Over the past few weeks, the Model Context Protocol (MCP) has rapidly emerged as the de facto standard for connecting third-party data and tools to LLM-based conversational systems. In a short time, the internet has been flooded with all kinds of flashy application cases, and the room for imagination is enormous. Yet behind this protocol there also lurk some subtle security risks and functional limitations.

Background

The essence of MCP is to act as a bridge for building an “intelligent assistant plugin ecosystem”, allowing users to provide local extension capabilities to various large models and AI clients that support the MCP protocol, so that these LLM-based intelligent interfaces can perform non-text operations through MCP’s “tool calls”. The MCP protocol in effect standardizes the way LLMs interface with local data and tools.

For example, in the previous article, we simply told the large model in Cursor: “Visit the www.saucedemo.com website, log in with problem_user, verify the complete workflow of the site, and record any issues found” , and Cursor then used the playwright-mcp MCP server to call the playwright framework to carry out a series of activities such as launching the browser, visiting, browsing, testing, and recording issues—completing testing activities similar to those of a manual test engineer

So through MCP, its core value is that it breaks through the original role of large AI models as text-based advisory assistants and turns them into actual executors, better able to directly operate local tools and data, providing large models with a universal socket.

The Potential Hidden Concerns

But while we recognize MCP’s enormous potential, we must also see that MCP, as a foundational protocol that has only recently gotten started, is far from perfect, and there are still the following hidden concerns:

Hidden Concern 1 - Security of the Protocol Itself

Authentication Mechanisms

The initial version of MCP did not define an authentication specification, leaving each MCP server to do as it pleased—from cumbersome verification to sensitive data being left completely exposed, a veritable chaos. Although authentication schemes were introduced later, their implementation complexity has sparked new controversy.

The Risk of Local Malicious Code Execution

The protocol supports running MCP servers via standard input/output, which, while lowering the barrier to use, also provides a breeding ground for the spread of malicious code. If a non-technical user runs third-party malicious MCP code and causes a local device security failure, this has become a new attack vector.

MCP’s Trust in User Input

Many MCP servers directly execute user input instructions. When user intent is translated by the LLM and takes effect on the system, traditional security models face a severe challenge: the boundary between user input instructions and attacks on the system is no longer clear.

Hidden Concern 2 - Limitations in Interaction Design

The MCP protocol is oriented more toward LLM interaction, but for users, the interaction is hardly easy to use or friendly.

The Lack of Risk Grading for Different Levels of Operations

Many assistant tools provide a global operation-confirmation mechanism, but the “auto-confirm” habit formed around many harmless tools often leads to serious misoperations, such as the deletion of important local files or sending emails by mistake.

Cost Control Capability

Unlike traditional protocols, in LLM scenarios different volumes of returned tokens incur varying costs, and the cost of serving users depends heavily on the token efficiency of MCP tools, yet the current protocol has no limit on return length to strengthen cost control.

Sacrificing the Advantages of Structured Data

To accommodate LLM preferences, MCP tends to adopt human-readable formats such as plain text / images / audio clips for transmission, sacrificing the advantages of structured data. For services such as maps or Didi-style ride-hailing, which need to ensure location accuracy, trip-detail feedback, and real-time status updates, the current protocol is hardly up to the task. Solving such problems for now still requires some workaround designs at the tool layer.


Hidden Concern 3 - LLM Security

The discussion around the security of large models themselves is still unresolved, and introducing MCP to connect more data and tools makes this problem even more prominent.

MCP Makes Prompt Injection More Powerful

LLMs usually have two layers of instructions: system-level prompts (which control the assistant’s behavior and policies) and user-level prompts (provided by the user). Usually, when prompt injection was mentioned in the past, it referred to malicious user input. A fairly significant vulnerability of the MCP model is that it allows tools provided by third parties to be trusted as part of the assistant’s system-level prompt, thereby granting them greater privileges to override the agent’s behavior.

Sensitive Data Leakage

Beyond malicious hijacking via malicious tools, MCP can also cause users to leak sensitive data unintentionally. Even if a user strictly confirms every operation, data may still be exfiltrated in a “reasonable” way. For example, with your instruction: “Explain my personal medical records to me”, the assistant may launch a medical-retrieval multi-tenant MCP tool and leak your private personal information to the service provider.

MCP servers can also easily hijack or alter user intent through disguised prompts and their actual tool implementations.

Hidden Concern 4 - Breaking Traditional Enterprise Data Access Controls

Unlike the direct leakage of sensitive data, after an enterprise connects large amounts of internal data to an LLM via MCP, ordinary employees may be able to access or infer data they are not authorized to access.

For example, if all of an enterprise’s internal documents are connected to an LLM, an employee, through the assistant:

“Based on all the documents I have permission to access and the company’s appointment and removal notices, infer major events that have not yet been announced, such as layoffs”

Or someone at the manager level:

“I saw an anonymous negative comment about XXX on the internal wiki, the content is XXXX; please analyze the speaking records of members in the XX team and infer who most likely wrote this comment”

Hidden Concern 5 - The Inherent Capability Limits of LLMs

For large-scale MCP integration, the load capacity of the LLM itself may well be unable to cope. Google’s A2A protocol (Agent to Agent) may be a solution, but that is another matter.

MCP’s Effectiveness Depends on the LLM Integrated into the Assistant

One fact is that it is not the case that the more tools and the more data, the more accurate the results; on the contrary, the more similar data and tools there are, the worse the precision of the final output becomes. As the number of MCP servers grows and more tools are integrated, the performance of AI assistants will also decline significantly, and the cost of requests and analysis will increase.

The Gap Between Tool Implementation and User Expectations

In implementation, many MCP servers are essentially functions or methods written to call tools that already exist locally, so their capabilities are mainly constrained by those functions and methods. If users have little understanding of the specific function implementations and simply interact with the LLM following ordinary human communication patterns, they are likely to struggle to get satisfactory results.

“Please find for me all recent blog articles related to AI testing”; this instruction actually requires the local MCP server to have the ability to full-text search local text content, but if the integrated MCP server only includes file-finding functionality, naturally no corresponding results can be obtained.

Conclusion and Outlook

The birth of MCP answers the data-integration needs of the AI era, and its everyday application value is beyond doubt. But we also need a sober awareness: the MCP protocol is still in its growth phase, and while we embrace its applications, we must also recognize the risks within it; combining LLMs with data amplifies existing risks and gives rise to new types of threats.


Most of the views in this article are drawn from Shrivu’s blog (Read the original article)

Built with Hugo
Theme Stack designed by Jimmy