MCP risk scores in Security Center: what an admin needs to know
The short answer: For MCP Salesforce security, Winter ’27 adds a first filter. Salesforce scans each third-party MCP server when it’s registered, gives it a Low, Medium or High risk score in Security Center, keeps monitoring it for changes, and lets admins block high-risk servers. It doesn’t tell you what a server will do with your data, so permissions and a short list of questions still decide the outcome.
A developer on your team messages you at 4:30 on a Thursday. They found an MCP server that lets an AI assistant look up Salesforce accounts, update a field and draft a follow-up. It has a tidy README and a few hundred stars. “Can we connect it for a test?” You don’t know who runs it, what it logs, or what happens to the account notes it reads. You do know the assistant will act as a real user.
Winter ’27 gives you a first signal for that conversation. It isn’t the whole answer. Features vary by edition and license, so check the release notes for your org before you rely on any of this.
What is MCP, in two sentences?
MCP, the Model Context Protocol, is an open standard for connecting AI applications to outside data sources, tools and workflows; its own documentation compares it to a USB-C port for AI. An MCP server is the thing on the other end of that port: it lists the actions an AI client may call (look up a record, update a field), and the client calls them on someone’s behalf.
For an admin, the practical point is that an MCP connection is an integration with a new kind of caller. The caller is software that decides at run time which tool to use, and it runs with whatever access the connecting user has.
What do MCP risk scores do in Security Center?
Per Salesforce’s Winter ’27 platform blog (October 13, 2026), Security Center assigns each third-party MCP server a Low, Medium or High risk score. Servers are scanned at registration and then monitored for changes, and admins can set policies and block high-risk servers without investigating each one by hand. The blog says all its features are generally available unless marked otherwise, and only Observability Center is marked beta.
Salesforce’s Winter ’27 Trailhead module lists the feature under a “Salesforce Guardian” heading and gives the reason: third-party MCP servers “can be exposed to tool poisoning and rug pull attacks”, and the scores are meant to “proactively identify server threats before they compromise agent workflows.” For where Guardian sits among the Dreamforce names, see our plain-English guide to the new names.
What don’t the sources tell us?
Quite a lot, and it matters before you build a process on the score. In the sources we could open, Salesforce doesn’t say how the score is calculated, which signals feed it, whether an admin can override it, or which Security Center edition and license you need. Confirm what your org has, in Setup and with your account team.
Treat the score as a screening step. It’s a signal about the server, not an assessment of your data, your permissions or your use case.
What are tool poisoning and rug pulls?
Tool poisoning hides instructions in a tool’s description that the AI follows, though a human reading the tool list would never notice. Salesforce’s January 2026 Agentforce MCP post describes it as a form of indirect prompt injection and recommends vetting each server and tool closely. A rug pull, in our plain-English use of the term, is a server that behaves while you evaluate it and then changes its tools or behavior later.
That second one is why “monitored for changes” is the useful half of the feature. A review at registration only tells you about registration day.
What should you ask before an MCP server reaches org data?
Ask these, in this order, and write the answers down. Salesforce’s docs don’t prescribe them; they’re our opinion of a sensible minimum.
- Who runs it? A named company or person, a contact, and where it’s hosted.
- What tools does it expose? Read, write or delete? Salesforce’s own post says admins can choose which tools to import from a server, with each tool becoming an action. Import the smallest set that does the job.
- What leaves Salesforce? Which fields and records does it receive, where are they stored, and for how long?
- What does the risk score say, and when did it last change? A Medium that was Low last month deserves a conversation.
- Who is the connecting user? See the next section. This is where most of the risk sits.
- How will you see what it did? Decide before connecting how you’ll find out what the server changed.
- How do you turn it off? Name the person who can, and how fast.
- Who re-reviews it, and when? A server you approved in November isn’t the same server in May.
How do you give the connecting user least privilege?
Give the connecting user the narrowest access that lets the use case work, because the server can do whatever that user can do. Salesforce’s hosted MCP documentation describes requests running as the authenticated user, set up through an External Client App, and says you can restrict who may authenticate by requiring a permission set. Confirm the details against the current docs, because the product is still moving.
In practice:
- Create a dedicated permission set for the MCP use case. Start from nothing and add the objects and fields the use case reads. Add edit access only when read-only has worked for a while.
- Don’t connect with an administrator. An assistant running as a System Administrator, or any user with View All Data or Modify All Data, can reach everything. Use an ordinary user’s access, or a purpose-built integration user.
- Use one External Client App per client type, and require the permission set for pre-authorization so only approved people can connect.
- Check field-level security, not only object access. A server that can read Contact can read every Contact field the user can.
- Try it in a sandbox first, as an ordinary user, and read what comes back. Look for the field you hoped was hidden.
Nothing here needs a new product. Permission sets, field-level security and the Security Center scores are already in the platform. Our guide to running a user access review covers the permission-reading part, and how to tell what an AI agent changed covers the afterwards.
What to do Monday:
– Open Security Center and find any third-party MCP servers already registered. Note each score and who registered it.
– For each server, name an owner and write down the connecting user and their permission sets.
– Block anything scored High unless someone can say in one sentence why it’s needed.
– Draft the eight questions above as a one-page form that anyone requesting a new connection must fill in.
Where we come in
Our Agentforce Readiness Sprint looks at your data and permissions through an agent’s eyes: what the connecting user can see and change, usually more than anyone intended. It ends with a plan, a guardrail design and a pilot with a stop condition. It doesn’t build the integration, and it isn’t a licensing review.
Questions people ask
Does a Low risk score mean an MCP server is safe?
No. Low means the scan found less to worry about at the time. It doesn’t describe what the server does with your data once access is granted, and it can change. You still need permissions, a named owner and the questions above. Treat Low as “fine to look at”, not “fine to approve”.
Does the risk score apply to Salesforce’s own MCP servers?
The sources we found describe scores for third-party servers. They don’t say how Salesforce-hosted servers are treated, so don’t assume. Hosted servers run as the authenticated user, so least privilege applies to them in the same way. Check the documentation for your release.
Can I block an MCP server in Salesforce?
According to Salesforce’s Winter ’27 platform blog, yes: admins can set policies and block high-risk servers in Security Center. Separately, admins register third-party servers in Setup and choose which tools to import, so an unregistered server shouldn’t be reachable by your agents. Test that assumption in a sandbox before relying on it.
Sources
- Platform Winter ’27 release: MCP Risk Scores, scanning, blocking and the GA statement.
- Trailhead: Winter ’27 highlights, Headless 360 and Guardian: the Guardian heading, tool poisoning and rug pull wording.
- Salesforce: Agentforce MCP support: registering servers in Setup, the tool allowlist and tool poisoning (January 2026, beta at the time).
- Model Context Protocol: introduction: what MCP is.
- Salesforce developer docs: create an External Client App for hosted MCP servers: External Client App setup and permission set restriction.