MCP security checklist: audit, harden and govern the MCP servers your staff use
MCP servers give AI agents hands: they read mail, query databases and run commands with real tokens. Most companies have more of them than they think, and none were reviewed. Work through these checks in order.
Eight areas, 40 checks
Next step
AI Agent Blocklist: the MCP part of your checklist, done daily
MCP server domains, MCP packages and MCP bridge extensions in one list for firewalls, DNS filters and device policies.
From $99 a month for one organisation
Why MCP needs its own security checklist
The Model Context Protocol is the plug that connects an AI agent to tools and data. It is simple to install, which is why it spread so fast, and it sits outside most existing reviews. Six problems come up in almost every MCP security audit.
Shadow servers
Developers add servers from directories in seconds. Nobody in IT knows they exist or what they can reach.
Broad tokens
Servers run with personal access tokens or service keys that can read far more than the task needs.
Hidden instructions
Text in a tool description, a ticket or a web page can steer the agent to read files or send data.
Supply chain
A server is a package. A new version can change behaviour after you approved the old one.
Local code execution
Local servers start as processes on the laptop with the user's rights and reach whatever the user can reach.
No audit trail
Tool calls often happen without a log you can search, so an incident cannot be traced afterwards.
The 2025 patterns behind the checks
Each check below answers a pattern that public research showed in 2025. We describe the mechanism only, at the level a defender needs, and never the products involved.
- 2025
Tool poisoning named. Researchers showed that hidden text in a tool description can instruct the agent, and that a server can change its descriptions after the user approved them.
- 2025
Private data exposed through an issue tracker. A planted issue asked an agent to read private repositories through an MCP server and publish them. The token was broader than the task needed.
- 2025
Developer tools that could run commands. Some MCP debugging tools and client proxy packages could be made to run commands by a request or server from outside. Keep such tools updated and bound to the local machine.
- 2025
Database secrets copied into a support ticket. An agent with a broad database role read hidden instructions in a ticket and pasted private token data into the thread.
- 2025
Servers open on the internet. Researchers found MCP servers reachable from the internet that listed their tools to anyone, without any authentication. Check that every server you run requires sign-in.
- 2025
A look-alike server on a package registry. A mail server package added one line in a new version that copied every outgoing email to an outside address. Pinned versions and update reviews catch this.
The MCP security checklist
Forty checks in eight groups. Work top to bottom: you cannot approve what you have not found, and you cannot audit what you do not log. Copy it into your own wiki or ticket template.
Inventory: find every MCP server
An MCP security audit starts here- Search managed devices for MCP config files of each AI client in use, such as .mcp.json in projects
- List installed MCP packages and CLIs, for example @smithery/cli or context7-mcp
- Find MCP bridge extensions in browsers, such as Browser MCP bjfgambnhccakkhmkepdoekmckoijdlc
- Pull DNS and proxy logs for hosted MCP endpoints and MCP directories
- Ask each team lead which connectors their AI accounts use, then compare answers with the logs
Approval and allowlist
Only named servers, each with an owner- Keep one approved list with server name, version, source, owner and review date
- Enforce it through each client's managed settings, not a wiki page (see our MCP allowlist guide)
- Turn off automatic approval of project-level server files in coding tools
- Block MCP directories and hosted servers that are not on the list at the network
- Write down who can approve a new server and how fast they answer
Identity and tokens
The token decides the damage- Give each server its own token with the smallest scope that works
- No personal admin tokens in MCP configs, and no tokens in shared project files
- Prefer remote servers that use OAuth sign-in over pasted long-lived keys
- Check that servers never pass the user's token through to other services
- Set token expiry and rotate after every staff change in the owning team
Transport and network
Know what crosses the wire- Self-hosted HTTP servers listen on localhost or behind authentication, never open
- Require TLS for every remote server
- Retire servers still using the deprecated HTTP plus SSE transport
- Block unapproved MCP domains at the firewall or DNS filter (see AI at the firewall)
- Limit which internal hosts a server can reach, to avoid request forgery into your network
Tool review
Read what the model reads- Read every tool description before approval and look for instructions aimed at the model
- Pin server versions and re-review on every update
- Disable tools you do not need, especially write, delete and send tools
- Check for tool names that shadow tools of another server
- Prefer servers from the vendor of the system they connect to
Runtime limits
Contain what a server can do- Keep human approval on for tools that write, send, pay or delete
- Run local servers in a container or sandbox where the client supports it
- Never give a server access to the home folder, SSH keys or cloud credentials
- Separate servers that read untrusted content from servers that can send data out
- Block coding agents' "skip all permissions" modes in managed settings
Logging and audit
Make tool calls searchable- Log every tool call with user, server, tool, time and result
- Send client and gateway logs to your SIEM
- Alert on new MCP domains and new MCP packages on devices
- Review the approved list every quarter and remove unused servers
- Run the inventory again each month and diff it against the list
Incident response
Have the steps ready- A runbook to revoke all tokens of one server in minutes
- A way to remove a server from every device through MDM
- A network block you can push for one MCP domain right away
- Contacts at the server's maintainer or vendor
- A list of data each server could reach, ready for a breach assessment
MCP security hardening: the managed setting in each client
Checks 2 and 6 depend on settings users cannot change. Each AI client has its own. This table lists the controls we verified in each vendor's documentation; menus change, so check the current admin console before you roll out.
| Client | Allow only approved servers | Turn MCP off | Where it is enforced |
|---|---|---|---|
| Claude Code | allowedMcpServers and deniedMcpServers, matched by name, command or URL | managed-mcp.json with no servers | System folder, device management, or the Claude admin console |
| Codex | mcp_servers entries with an identity (command or URL) | Empty mcp_servers section | requirements.toml or macOS device management |
| VS Code with Copilot | ChatAllowedMcpServers policy, or GitHub's MCP allowlist | ChatMCP policy set to none | Group Policy, macOS profile, Linux policy file, GitHub settings |
| Cursor | Team MCP allowlist with commands and URLs | Allowlist with no entries | Team settings on the Enterprise plan |
| Gemini CLI | mcp.allowed and mcp.excluded | Admin MCP setting in system settings | System settings file that overrides users |
| Claude Desktop | Desktop extension allowlist | isLocalDevMcpEnabled set to false | Device policy or organisation settings |
| JetBrains AI Assistant | MCP servers chosen per profile | Users not allowed to add servers | IDE Services AI Enterprise settings |
The specification's security topics, turned into admin actions
The MCP specification, first published by Anthropic in November 2024, has a Security Best Practices page. It is written for people who build servers and clients. This table maps its topics to what an admin can check.
| Topic in the specification | What it means | What you check |
|---|---|---|
| Confused deputy problem | A proxy server reuses its own consent to act for a different client | Proxy servers ask for consent per client; no shared approvals |
| Token passthrough | A server forwards a token it was not issued for | Servers you approve validate tokens and never pass them on |
| Server-side request forgery | A malicious server makes the client fetch internal addresses | Clients and gateways cannot reach internal metadata or admin hosts |
| Session or state handle hijacking | Someone reuses an identifier to inject into another user's work | Servers bind identifiers to the user and use random values |
| Local MCP server compromise | A local server runs harmful commands on the device | Local servers need explicit approval and run with limited rights |
| Authorization URL validation | A sign-in URL from a server is used to run code or phish | Clients are on versions that validate sign-in URLs |
| Scope minimisation | Servers ask for broad scopes up front | Tokens start narrow and grow only on request |
MCP security policy: clauses to put in writing
A policy gives your controls a reason people can read. These clauses fit inside an existing acceptable use policy. Our AI agent acceptable use policy page has the wider template.
Approved servers only
Staff may connect AI tools only to MCP servers on the approved list. Requests go to the named approver.
Enforced by managed settingsNo secrets in configs
Tokens and keys for MCP servers live in the secret store, never in project files or chat.
Checked in code reviewLeast privilege
Each server gets its own token with the smallest scope that works, with an expiry date.
Checked at approvalHuman approval for actions
Tools that send, pay, delete or publish need a person to confirm each call.
Enforced by client settingsLogging
Tool calls are logged centrally and kept for the same period as other security logs.
Checked each quarterReview and removal
Approved servers are reviewed each quarter and on every version change. Unused servers are removed.
Owned by the server ownerMCP security governance: who owns what
MCP touches engineering, IT and security at once, so checks fall between teams. Name an owner for each area before the first audit.
Security
- Owns the policy and the approved list
- Reviews new servers
- Runs the quarterly audit
IT and endpoint
- Pushes managed settings
- Runs device inventory
- Removes unapproved servers
Network
- Loads the block list
- Watches DNS and proxy logs
- Pushes emergency blocks
Engineering
- Proposes servers with a reason
- Keeps tokens narrow
- Owns each approved server
How we help with the blocking part of the checklist
Several checks need a list of what MCP looks like from the outside: domains, packages and extensions. That list changes every day, and it is the part we maintain for you.
Network
- Domains of hosted MCP servers and integrations
- MCP directories such as smithery.ai and glama.ai
- Firewall, DNS filter, hosts and PAC formats
Devices
- MCP packages and CLIs with their package names
- Process names to search for in inventory
- Desktop agent apps with bundle and package IDs
Browsers
- Extension IDs of MCP bridge and agent extensions
- Ready block lists for Chrome and Edge policies
- Risk grade on every entry, Low to Severe
Keep the MCP checks current without the manual work
New MCP servers, packages and bridge extensions appear every day. The AI Agent Blocklist adds them as they appear, and we set up your download access within one business day of your order.
- 2,600+ AI agent domains, MCP servers included
- 5,300+ identifiers: extensions, apps, packages, processes
- CSV and JSON with category and risk grade
- Firewall, DNS, hosts, PAC and browser policy files
- Updated every day
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
MCP security questions
What should an MCP security audit cover?
Four things: which MCP servers exist on your devices and in your accounts, which of them are approved, what each server can reach with its tokens, and whether tool calls are logged.
Start with an inventory from devices and network logs, then compare it with your approved list.
What are the MCP security best practices in the specification?
The MCP specification has a Security Best Practices page. Its topics include the confused deputy problem, token passthrough, server-side request forgery, session or state handle hijacking, local server compromise, authorization URL validation and scope minimisation.
Most of it is aimed at people who build servers and clients. The table above turns it into admin checks.
What is MCP tool poisoning?
Tool poisoning hides instructions in a tool's description. The model reads the description, the user usually does not, and the hidden text can tell the agent to read files or send data.
Researchers named it in April 2025. Review tool descriptions before approval and pin versions so a description cannot change silently.
Should we allow remote MCP servers or only local ones?
Both carry risk. Local servers run code on the laptop with the user's rights. Remote servers receive data and tokens over the internet.
Allow only named servers of either kind, and block unapproved hosted MCP endpoints and directories at the firewall or DNS filter.
How do we stop staff adding unapproved MCP servers?
Use the managed settings of each MCP client to allow only approved servers, block hosted MCP domains at the network, and block MCP bridge extensions in the browser. Check devices for MCP packages and config files each month.
Can we put this MCP security checklist in our own wiki or GitHub repository?
Yes. Copy the checks into your internal wiki, repository or ticket template and adapt the owners and review dates to your teams.
How often should approved MCP servers be reviewed?
At least every quarter and on every version change. Re-check the tool descriptions, the token scopes and the owner. Remove servers nobody has used for 90 days.
Anthropic, Claude, Claude Code and Claude Desktop are trademarks of Anthropic, PBC. Codex is a trademark of OpenAI. Cursor is a trademark of Anysphere, Inc. Gemini and Chrome are trademarks of Google LLC. GitHub and GitHub Copilot are trademarks of GitHub, Inc.
Microsoft, VS Code and Edge are trademarks of Microsoft Corporation. JetBrains is a trademark of JetBrains s.r.o. Smithery, Glama, Browser MCP and Context7 are trademarks of their respective owners. OWASP is a trademark of the OWASP Foundation.
Names are used only to identify the products discussed. AI Agent Blocklist is not affiliated with, sponsored or endorsed by these companies.