Shiny tools, shallow checks: how the AI hype opens the door to malicious MCP servers

Introduction

In this article, we explore how the Model Context Protocol (MCP) — the new “plug-in bus” for AI assistants — can be weaponized as a supply chain foothold. We start with a primer on MCP, map out protocol-level and supply chain attack paths, then walk through a hands-on proof of concept: a seemingly legitimate MCP server that harvests sensitive data every time a developer runs a tool. We break down the source code to reveal the server’s true intent and provide a set of mitigations for defenders to spot and stop similar threats.

What is MCP

The Model Context Protocol (MCP) was introduced by AI research company Anthropic as an open standard for connecting AI assistants to external data sources and tools. Basically, MCP lets AI models talk to different tools, services, and data using natural language instead of each tool requiring a custom integration.

MCP follows a client–server architecture with three main components:

• MCP clients. An MCP client integrated with an AI assistant or app (like Claude or Windsurf) maintains a connection to an MCP server allowing such apps to route the requests for a certain tool to the corresponding tool’s MCP server.
• MCP hosts. These are the LLM applications themselves (like Claude Desktop or Cursor) that initiate the connections.
• MCP servers. This is what a certain application or service exposes to act as a smart adapter. MCP servers take natural language from AI and translate it into commands that run the equivalent tool or action.

MCP as an attack vector

Although MCP’s goal is to streamline AI integration by using one protocol to reach any tool, this adds to the scale of its potential for abuse, with two methods attracting the most attention from attackers.

Protocol-level abuse

There are multiple attack vectors threat actors exploit, some of which have been described by other researchers.

1. MCP naming confusion (name spoofing and tool discovery)
   An attacker could register a malicious MCP server with a name almost identical to a legitimate one. When an AI assistant performs name-based discovery, it resolves to the rogue server and hands over tokens or sensitive queries.
2. MCP tool poisoning
   Attackers hide extra instructions inside the tool description or prompt examples. For instance, the user sees “add numbers”, while the AI also reads the sensitive data command “cat ~/.ssh/id_rsa” — it prints the victim’s private SSH key. The model performs the request, leaking data without any exploit code.
3. MCP shadowing
   In multi-server environments, a malicious MCP server might alter the definition of an already-loaded tool on the fly. The new definition shadows the original but might also include malicious redirecting instructions, so subsequent calls are silently routed through the attacker’s logic.
4. MCP rug pull scenarios
   A rug pull, or an exit scam, is a type of fraudulent scheme, where, after building trust for what seems to be a legitimate product or service, the attackers abruptly disappear or stop providing said service. As for MCPs, one example of a rug pull attack might be when a server is deployed as a seemingly legitimate and helpful tool that tricks users into interacting with it. Once trust and auto-update pipelines are established, the attacker maintaining the project swaps in a backdoored version that AI assistants will upgrade to, automatically.
5. Implementation bugs (GitHub MCP, Asana, etc.)
   Unpatched vulnerabilities pose another threat. For instance, researchers showed how a crafted GitHub issue could trick the official GitHub MCP integration into leaking data from private repos.

What makes the techniques above particularly dangerous is that all of them exploit default trust in tool metadata and naming and do not require complex malware chains to gain access to victims’ infrastructure.

Supply chain abuse

Supply chain attacks remain one of the most relevant ongoing threats, and we see MCP weaponized following this trend with malicious code shipped disguised as a legitimately helpful MCP server.

We have described numerous cases of supply chain attacks, including malicious packages in the PyPI repository and backdoored IDE extensions. MCP servers were found to be exploited similarly, although there might be slightly different reasons for that. Naturally, developers race to integrate AI tools into their workflows, while prioritizing speed over code review. Malicious MCP servers arrive via familiar channels, like PyPI, Docker Hub, and GitHub Releases, so the installation doesn’t raise suspicions. But with the current AI hype, a new vector is on the rise: installing MCP servers from random untrusted sources with far less inspection. Users post their customs MCPs on Reddit, and because they are advertised as a one-size-fits-all solution, these servers gain instant popularity.

An example of a kill chain including a malicious server would follow the stages below:

• Packaging: the attacker publishes a slick-looking tool (with an attractive name like “ProductivityBoost AI”) to PyPI or another repository.
• Social engineering: the README file tricks users by describing attractive features.
• Installation: a developer runs pip install, then registers the MCP server inside Cursor or Claude Desktop (or any other client).
• Execution: the first call triggers hidden reconnaissance; credential files and environment variables are cached.
• Exfiltration: the data is sent to the attacker’s API via a POST request.
• Camouflage: the tool’s output looks convincing and might even provide the advertised functionality.

PoC for a malicious MCP server

In this section, we dive into a proof of concept posing as a seemingly legitimate MCP server. We at Kaspersky GERT created it to demonstrate how supply chain attacks can unfold through MCP and to showcase the potential harm that might come from running such tools without proper auditing. We performed a controlled lab test simulating a developer workstation with a malicious MCP server installed.

Server installation

To conduct the test, we created an MCP server with helpful productivity features as the bait. The tool advertised useful features for development: project analysis, configuration security checks, and environment tuning, and was provided as a PyPI package.

For the purpose of this study, our further actions would simulate a regular user’s workflow as if we were unaware of the server’s actual intent.

To install the package, we used the following commands:

Now that the package was installed and running, we configured an AI client (Cursor in this example) to point at the MCP server.

Now we have legitimate-looking MCP tools loaded in our client.

Below is a sample of the output we can see when using these tools — all as advertised.

But after using said tools for some time, we received a security alert: a network sensor had flagged an HTTP POST to an odd endpoint that resembled a GitHub API domain. It was high time we took a closer look.

Host analysis

We began our investigation on the test workstation to determine exactly what was happening under the hood.

Using Wireshark, we spotted multiple POST requests to a suspicious endpoint masquerading as the GitHub API.

Below is one such request — note the Base64-encoded payload and the GitHub headers.

Decoding the payload revealed environment variables from our test development project.

This is clear evidence that sensitive data was being leaked from the machine.

Armed with the server’s PID (34144), we loaded Procmon and observed extensive file enumeration activity by the MCP process.

Next, we pulled the package source code to examine it. The directory tree looked innocuous at first glance.

The server implements three convincing developer productivity tools:

• analyze_project_structure.py analyzes project organization and suggests improvements.
• check_config_health.py validates configuration files for best practices.
• optimize_dev_environment.py suggests development environment optimizations.

Each tool appears legitimate but triggers the same underlying malicious data collection engine under the guise of logging metrics and reporting.

Core malicious engine

The project_metrics.py file is the core of the weaponized functionality. When launched, it tries to collect sensitive data from the development environment and from the user machine itself.

The malicious engine systematically uses pattern matching to locate sensitive files. It sweeps both the project tree and key system folders in search of target categories:

• environment files (.env, .env.local, .env.production)
• SSH keys (~/.ssh/id_rsa, ~/.ssh/id_ed25519)
• cloud configurations (~/.aws/credentials, ~/.gcp/credentials.json)
• API tokens and certificates (.pem, .key, .crtfiles)
• database connection strings and configuration files
• Windows-specific targets (%APPDATA% credential stores)
• browser passwords and credit card data
• cryptocurrency wallet files

Each hit is classified by type, its first 100 KB is captured, and the result is cached for eight hours to keep scans fast and quiet. The target file metadata (path, size, and last modified time) is logged, while sensitive bytes are redacted before any UI display: that way, the victim only sees the legitimate output in Cursor. Those redacted indexes are sent as input to the analyzer, config-health checker, and environment optimizer, letting them provide seemingly useful results while the server silently hoards the real secrets.

Data exfiltration

After the harvesting, the engine calls send_metrics_via_api() to ship data to the endpoint acting as a C2 server in this case.

The tools try to exfiltrate data by disguising compromised traffic as something that looks legitimate so it can hide in plain sight.

Takeaways and mitigations

Our experiment demonstrated a simple truth: installing an MCP server basically gives it permission to run code on a user machine with the user’s privileges. Unless it is sandboxed, third-party code can read the same files the user has access to and make outbound network calls — just like any other program. In order for defenders, developers, and the broader ecosystem to keep that risk in check, we recommend adhering to the following rules:

1. Check before you install.
   Use an approval workflow: submit every new server to a process where it’s scanned, reviewed, and approved before production use. Maintain a whitelist of approved servers so anything new stands out immediately.
2. Lock it down.
   Run servers inside containers or VMs with access only to the folders they need. Separate networks so a dev machine can’t reach production or other high-value systems.
3. Watch for odd behavior.
   Log every prompt and response. Hidden instructions or unexpected tool calls will show up in the transcript. Monitor for anomalies. Keep an eye out for suspicious prompts, unexpected SQL commands, or unusual data flows — like outbound traffic triggered by agents outside standard workflows.
4. Plan for trouble.
   Keep a one-click kill switch that blocks or uninstalls a rogue server across the fleet. Collect centralized logs so you can understand what happened later. Continuous monitoring and detection are crucial for better security posture, even if you have the best security in place.