MCP security risks: why MCP servers need controls

MCP lets AI agents call tools and read data through servers that anyone can publish. Each server is code or a service running with your users' access, and the model trusts what those servers tell it. Without controls, that is a new and wide open door.

Next step

AI Agent Blocklist: MCP server packages, processes and domains, so unapproved servers cannot be installed or reached

Part of one list of AI agents, coding agents and agent tools, with a risk grade on every entry and daily updates.

From $99 a month for one organisation See plans Ask a question

Is MCP a security risk?

The protocol is not the problem; unmanaged use is. The MCP specification defines an authorization model for remote servers and publishes security guidance. Those protect you only if servers and clients follow them.

The risk comes from three facts: servers are third-party code or services, local servers run with the user's rights, and the AI model reads tool descriptions and results that an attacker can shape.

Local servers (stdio)

  • Started as a process on the user's machine
  • Often installed straight from npm or PyPI on first use
  • Can run any command they are configured with
  • Reach files, keys and the local network

Remote servers (HTTP)

  • A service run by the vendor or a third party
  • Receive company data and access tokens
  • Depend on correct OAuth handling
  • Can change behaviour at any time on their side

MCP security threats, explained for admins

These are the threats described in public research and in the MCP specification's own security guidance. The tags show where each one happens.

Tool poisoning

Hidden instructions in a tool's description. Users rarely see descriptions; the model always does.

ModelLocalRemote

Rug pull

A server you approved changes its tools or descriptions later, so the trusted server becomes a harmful one.

LocalRemote

Shadowing

One server's description changes how the agent uses another server's tools, without being called itself.

Model

Injection through results

Data a tool returns, such as an issue, email or web page, contains instructions the agent follows.

Model

Malicious packages

A server package built or modified to copy data or run hidden commands, published under a familiar name.

Local

Local command execution

A startup command in a client configuration can run anything. One-click install flows make this easy to miss.

Local

Confused deputy

A proxy server with a static client ID can let an attacker obtain authorization without the user's consent.

Remote

Token passthrough

A server accepts tokens not issued to it and passes them on, bypassing controls and audit trails. The specification forbids it.

Remote

Over-broad scopes

Tokens granted every scope up front give anyone who obtains one access to everything at once.

Remote
The specification also warns about server-side request forgery during OAuth discovery, session hijacking on multi-server deployments, and malicious authorization URLs that can lead to code execution in clients. Those are mostly for builders of MCP clients and servers, but they explain why patching clients matters.

MCP risk patterns reported in 2025

Public research in 2025 showed a handful of patterns again and again. We describe the patterns, not products; each one maps to a control further down this page.

Pattern 1

Tool poisoning demonstrated

Researchers showed a harmless-looking tool whose hidden description told the agent to read SSH keys and configuration files and send them to the server, while the agent's explanation talked only about the visible task.

Pattern 2

Public content, private access

An agent that reads public content, such as an issue, while it holds a token for private repositories can be steered to copy private data into a public place. Narrow tokens prevent this.

Pattern 3

Developer tools without authentication

Some tools for testing MCP servers listened on the network without authentication, so a request from outside could launch commands. Keep such tools updated and bound to the local machine.

Pattern 4

Command injection in connection helpers

Some helper packages for connecting AI apps to remote MCP servers could be made to run commands by an untrusted server's authorization response. Connect only to servers you have reviewed.

Pattern 5

Injected MCP configs

Prompt injection in AI code editors could create a new MCP configuration file, or swap an approved one, and run commands without a fresh prompt. Treat config files as sensitive and review every change.

Pattern 6

A look-alike email server

A server package for sending email added a hidden blind copy of every outgoing message to an outside address after several clean releases. Pinned versions and update reviews catch this.

Common claims about MCP security: what is true and what is not

Forum threads are often harsh about MCP. Some of that is fair criticism of how it has been deployed. Here is a plain sorting of the claims.

"MCP has no security model."
The specification defines authorization for remote servers and publishes security guidance. Many early servers and setups ignored them, which is a deployment problem.
"Local servers are harmless because they are local."
A local server is a program running with the user's rights. It can read the same files and keys the user can.
"If it is in a directory, someone reviewed it."
Listings check criteria, not code. One major AI vendor's documentation says it reviews connectors for listing but does not security-audit or manage any MCP server.
"I only use the tools I call."
Every connected server's descriptions reach the model. Shadowing works without the harmful tool being called.
"Approving a server once is enough."
Servers update. Pin versions for local servers and re-review remote ones, or a rug pull goes unnoticed.

Who is exposed to MCP risk

MCP started with developers, but it now ships in desktop AI apps and office tools too. Map who has it before you set rules.

Developers

Coding agents with MCP servers for code hosting, databases, cloud consoles and browsers.

Knowledge workers

Desktop AI apps connected to email, calendars, documents and chat through connectors.

Admins of connectors

People who add organisation-wide connectors that every user's assistant then reaches.

Signs of unmanaged MCP use

MCP servers rarely show up in software inventory, because most are started on demand by the AI client. Look for their configuration and their traffic instead.

On devices and in repositories

  • .mcp.json files committed to repositories, which give every clone the same servers
  • MCP entries in Claude Code's user configuration and in .cursor/mcp.json files
  • mcp_servers tables in Codex's config.toml
  • Server processes started by npx, uvx or similar runners from an AI client

In logs and admin consoles

  • Package proxy requests for MCP server packages you never approved
  • Outbound traffic from laptops to MCP services and integration platforms
  • Organisation-wide connectors added in AI app admin consoles
  • OAuth grants to new apps in your identity provider and SaaS tools

Questions to answer before approving an MCP server

Use these questions for every server a team requests. If you cannot answer one, the server waits.

QuestionGood answerWarning sign
Who publishes and maintains it?The vendor of the system it connects to, or your own teamAn unknown account, a fork, or a name close to a known vendor
What can its tools do?Read-only, or narrow writes the task needsRun commands, send messages or delete data by default
Which credentials does it need?Its own scoped, short-lived tokenA personal admin token or a broad API key
Where does data go?Only to the system it connects toA third-party service in between
How do updates arrive?Pinned versions you upgrade on purposeLatest version pulled on every start
Credentials are the heart of it. Our page on AI agent identity and permissions covers scoped credentials for agents, and keeping agents away from secrets covers the developer machine.

Remote connectors in AI apps are MCP too

Many AI apps now offer connectors to email, files and work tools. Several of them are MCP servers behind a friendly name, and an admin can switch them on for everyone at once.

Organisation-wide reach

A connector added by one admin appears in every user's assistant, with each user's own access.

They follow users into tools

Claude Code, for example, fetches claude.ai connectors for signed-in users by default. Admins can switch that off with disableClaudeAiConnectors.

Same review applies

Treat each connector like a server request: who runs it, what it can do and which credentials it uses.

Other clients have their own switches: see Cursor's MCP Configuration and GitHub Copilot's MCP policy.

How to control MCP in the clients your staff use

Each AI client has its own MCP controls. Match servers by exact command or URL wherever you can; a server's name is a label a user chooses, not a security control.

ClientTurn MCP offAllow only approved servers
Claude Codemanaged-mcp.json with an empty mcpServers objectallowedMcpServers by serverCommand or serverUrl, with allowManagedMcpServersOnly; deniedMcpServers always wins; disableClaudeAiConnectors for claude.ai connectors
CodexAn mcp_servers table that is present but empty in requirements.tomlOne entry per approved server with an identity matching its command or URL
CursorMCP Configuration in the Enterprise dashboardAllowlist by command or URL, tool limits per server, network modes such as Allowlist or Deny all
GitHub CopilotThe MCP servers in Copilot policyAn MCP registry URL with Registry only, in public preview
VS CodeThe ChatMCP policyChatAllowedMcpServers, ChatDeniedMcpServers and ChatAllowManagedMcpServersOnly policies
Step-by-step instructions are on how to disable MCP servers and how to build an MCP allowlist. For coding agents specifically, see Claude Code and Codex CLI.

Six controls that cover most MCP risk

You do not need a new platform to start. These controls use settings you already have, plus a list of what to block.

  1. Inventory

    Find MCP configuration files on devices, connectors in your AI apps and server packages in your package proxy logs.

  2. Default deny

    Turn MCP off where nobody needs it, and allow only reviewed servers where they do, matched by command or URL.

  3. Pin and review

    Pin local server versions. Re-review when a version or a tool description changes.

  4. Least privilege

    Give each server its own scoped, short-lived credentials. Never reuse a user's broad token.

  5. Keep approvals on

    Do not let users set tool calls to always allow for servers that write, send or delete.

  6. Block the rest

    Block unapproved MCP packages and domains at the package proxy, on devices and on the network.

Logging helps you see what is really used. Claude Code, for example, can include MCP server and tool names in its OpenTelemetry events when OTEL_LOG_TOOL_DETAILS is set to 1. Check what each client you allow can export.

How the AI Agent Blocklist helps with MCP

Allowlists in each client cover the clients you manage. Our list covers the servers themselves, wherever someone tries to install or reach them.

What we listExamples from our fileWhere you block it
MCP server packages on npm and PyPI@browsermcp/mcp, @openai/codex-shell-tool-mcpInternal package proxies and registries
MCP server commands and processesfirecrawl-mcpEndpoint rules in Intune, Jamf or Defender
MCP services and integration platformsDomains in our MCP servers and integrations categoryFirewall, DNS filter, secure web gateway
Every entryTool name, category, identifier type and risk gradeBlock all, or the higher grades first
MCP is one part of the agent picture. Read AI coding agent security risks and AI agent security risks for the rest, then work through the MCP security checklist or go straight to blocking MCP servers.

Put a list behind your MCP rules

The AI Agent Blocklist includes MCP servers and integrations alongside AI agents, coding agents, AI browsers and agent extensions, updated every day.

From $99 a month for one organisation. Download access is set up within one business day of your order.

See plans Ask a question
  • MCP server package names for npm and PyPI
  • Process names and app identifiers for endpoint rules
  • 2,600+ AI agent domains, including MCP services
  • Plain, hosts, PAC and firewall list formats
  • CSV and JSON with category and risk grade

Ask us or order a plan

Tell us what you need to block and where your rules live. We reply within one business day.

  • Files for firewalls, DNS filters, secure web gateways, MDM and browser policies
  • Updated every day as new AI browsers, agents and extensions appear
  • Plans for one organisation, for MSPs and for product makers, or a hosted filter we run for you

We use your details only to answer you. No newsletters.

Questions about MCP security

Is MCP a security risk?
MCP itself is a protocol, and its specification includes an authorization model and security guidance. The risk comes from how it is used: servers are code or services from third parties, local servers run with the user's rights, and the AI model reads tool descriptions and results that attackers can shape. Unmanaged MCP use is a real risk.
What are the main MCP security issues?
Tool poisoning through hidden instructions in tool descriptions, servers that change after approval, malicious or compromised server packages, local servers that run arbitrary commands, prompt injection through data a tool returns, OAuth and token handling flaws such as confused deputy and token passthrough, and over-broad access scopes.
What is a malicious MCP server?
An MCP server built or modified to harm its users, for example by copying data to an outside address, running hidden commands, or feeding the model instructions. In 2025 a look-alike email server package was reported to copy outgoing email to an outside address after several clean releases.
Is MCP safe to use at work?
It can be, with controls. Many early setups ran any server a user found, with full access and no review, and the model can be steered by text in tool data. Those are deployment problems. With an allowlist, pinned versions, scoped credentials and approvals, MCP risk becomes manageable.
Are MCP servers listed in a directory safe?
Not necessarily. A listing means the server met the directory's criteria, not that it was security audited. At least one major AI vendor's documentation says it reviews connectors against listing criteria but does not security-audit or manage any MCP server.
How do I control which MCP servers employees use?
Use each client's admin controls to allow only approved servers by exact command or URL: managed-mcp.json and allowedMcpServers in Claude Code, mcp_servers requirements in Codex, MCP Configuration in Cursor, the MCP registry policy in GitHub Copilot and the MCP policies in VS Code. Block unapproved server packages and domains as a second layer.
Does turning off MCP break AI coding tools?
No. The assistants keep working with their built-in tools for reading code, editing files and running commands under your approval rules. What stops is access to extra tools and data sources. Most teams start with MCP off, then allow a short list of reviewed servers where there is a clear need.
Can an MCP server send data out without being called?
Researchers showed in 2025 that a harmful server's tool description can change how the agent uses a different, trusted server's tools, for example redirecting emails, without the harmful tool ever being called. This is why every connected server matters, not just the ones in use.

Claude, Claude Code and claude.ai are trademarks of Anthropic, PBC. Codex is a trademark of OpenAI. Cursor is a trademark of Anysphere, Inc. GitHub and GitHub Copilot are trademarks of GitHub, Inc.

VS Code, Intune and Defender are trademarks of Microsoft Corporation. Jamf is a trademark of Jamf. Browser MCP and Firecrawl are trademarks of their respective owners.

npm is a trademark of npm, Inc. PyPI is a trademark of the Python Software Foundation.

Names are used only to identify the products discussed. AI Agent Blocklist is not affiliated with, sponsored or endorsed by these companies.