Explore our new Cybersecurity Terms Glossary, explained by our experts.
Cybersecurity Terms Glossary
The Model Context Protocol standardises how AI assistants connect to tools and data. Securing it means treating every MCP server as a privileged integration rather than a plugin.

What Is Model Context Protocol Security?
The Model Context Protocol, released by Anthropic in November 2024 and subsequently adopted across the industry, gives AI assistants a standard way to connect to external tools and data sources. An MCP server exposes capabilities—query this database, read this directory, call this API—and any compliant client can use them. The appeal is obvious: one integration works everywhere. The security consequence is equally obvious once stated: a configuration file nobody reviews can grant a language model persistent access to production systems.
MCP defines the transport and the message format. It does not define an authorisation model, a trust model for servers, or a way for a client to verify that a server is what it claims to be. Those are left to implementers, and early implementations frequently skipped them. The result is a familiar pattern from the early days of browser extensions and mobile app permissions: a thriving ecosystem of community-published servers, installed with a copy-and-paste command, running with the privileges of the user who installed them.
Treat each server as a privileged integration subject to review, not as a convenience. Maintain an approved registry, pin versions, and verify the publisher and the artefact hash rather than installing from an arbitrary repository. Review the tool descriptions themselves, since they are untrusted text entering the model’s context, and re-review after an update because descriptions can change silently. Scope every credential to the minimum required and prefer short-lived tokens issued per session over a static key in a configuration file. Run servers in containers with no access to the wider network and no ambient credentials, and separate servers that reach sensitive data from those that reach the internet, so a single model session does not bridge the two. Require human approval for write operations, and log every tool call with its arguments. These are the same principles set out in our note on AI agent security, applied to a specific protocol.
Because MCP servers are typically installed by individual developers, the first governance question is discovery: find out which servers are running on managed devices and against which systems. From there, bring them into existing processes—an approved software list, change control for new connectors, and secret management rather than configuration files. Where a server reaches personal data, UK GDPR applies as it would to any integration, so the connector belongs in your record of processing and the access it grants belongs in your access review. For organisations in scope of DORA, an MCP server mediating access to a critical system is part of the ICT estate, not a developer tool. Protocol revisions have progressively added authorisation guidance, so track the specification version your clients and servers implement. At Privia Security, we review connector architectures, permissions, and server implementations through our AI Penetration Test Services.
No. It is a transport and message specification that deliberately leaves authorisation and trust to implementers, which means security depends entirely on how servers and clients are built and deployed. The risk comes from installing unreviewed community servers with broad, long-lived credentials and no sandbox, not from the protocol concept itself.
Check the publisher and whether the source is available; read the tool descriptions as untrusted input; identify what credentials and scopes it requires and whether they can be narrowed; confirm it can run containerised without host access; and pin a reviewed version with a verified hash. Re-check after updates, because tool descriptions and required scopes can change between releases.
A server that behaves correctly when a user first approves it, then changes its tool descriptions or behaviour in a later version to include malicious instructions. Because approval is usually granted once at installation, the change is invisible. Version pinning, hash verification, and re-review on update are the controls that address it.
Only against systems that do not matter, and ideally not at all on managed devices reaching corporate data. The practical compromise most organisations reach is an approved registry with a fast review path, plus discovery to find what is already installed. A slow approval process reliably produces unapproved connectors rather than fewer connectors.
You May Be Interested In These
We use a few cookies that are needed to make this website work, and they are always on. We would also like to set analytics cookies, to see which pages are useful; functional cookies, so that our live chat works; and advertising cookies, to measure our campaigns. We will not set any of those until you say yes, and you can change your mind at any time from the Cookie Settings link in the footer.
Choose the cookies you are happy for us to use. Strictly necessary cookies are always on; everything else is off until you turn it on.
These cookies keep the website working: they hold your security checks, your form submissions and the choice you make here. The site cannot run without them, so they are always on.
These let our live chat open and remember your conversation, and let the newsletter sign-up work. Switch them off and the chat window will not appear.
These tell us which pages are read and how people find us, so that we can improve the site. The reports are about groups of visitors, never about you personally.
These let us measure our advertising and show you our adverts on other websites, such as LinkedIn, Google and Meta. Switch them off and you will still see adverts, but they will be less relevant.