Skip to main content

Blackboard MCP Server: Security and AI Governance FAQ

Important

This document describes the Blackboard MCP Server at general availability and will be updated as the product changes. It is provided for information and does not modify any agreement between your institution and Blackboard. Nothing here is legal advice; consult your own counsel and procurement team on your obligations.

Here are the questions your security, privacy, and AI committees will ask before turning on the Blackboard® MCP Server, and the answers they need.

Item

Detail

Product

Blackboard MCP Server

Release

General availability with the December 2026 Blackboard LMS release (4001.7)

Version

1.0, November 2026

Audience

Information security, privacy, AI governance, and IT review committees

Endpoint

https://mcp.api.blackboard.com

The short version

Question

Short answer

Is it on by default?

No. A Blackboard LMS administrator must turn it on in the admin panel.

Can it reach data a user could not already reach?

No. Every call runs as the signed-in user, within that person's existing Blackboard LMS permissions.

Does the AI client see Blackboard LMS passwords?

No. Sign-in uses OAuth 2.0 on Blackboard and institution pages. The client receives only a time-limited token.

Does the server store our course or student data?

The server does not cache Blackboard LMS data. It keeps an audit log of each call, including the parameters sent, for 90 days.

Can it delete anything?

No. The tool catalog contains no delete tools.

Will our data train an AI model?

Not by Blackboard. Whether your AI vendor trains on it depends on the license tier you hold with that vendor.

Who decides when a person must approve an action?

Your institution, through the approval policy in your AI client. The server does not gate tool calls.

Can we see what was done?

Yes. Every call is logged, and you can request your institution's audit records through Blackboard Support.

What the server is, and what it is not

What is the Blackboard MCP Server?

The Model Context Protocol (MCP) is an open standard that lets an AI assistant call external tools in a consistent, structured way. The Blackboard MCP Server is Blackboard's hosted MCP endpoint for Blackboard LMS. It presents a curated catalog of tools, such as "list my courses" or "find students who have not submitted," that an approved AI assistant can call on a signed-in person's behalf.

It gives your institution one governed path for AI assistants to work with Blackboard LMS, with consistent authentication, tenant isolation, and auditing, instead of many separate integrations built and audited unevenly.

Why would we approve this instead of saying no?

Because the alternative to a governed path is usually an ungoverned one. Faculty and staff can already copy institutional data into AI assistants by hand or connect them through unofficial integrations. The MCP Server gives you a single place to switch access on or off, a defined set of actions, and an audit trail you can request.

Does it use any private or special access to Blackboard LMS?

No. The server is built on the public Blackboard LMS REST APIs, the same APIs available to any registered integration. It does not use private endpoints and gains no access beyond what those APIs and the user's own permissions allow.

Does the server include its own AI model?

No. The AI model belongs to the assistant your users choose, such as Claude, GitHub Copilot, or ChatGPT and Codex. The server receives a structured tool call from that assistant, runs it against Blackboard LMS as the signed-in user, and returns the result.

Which AI clients does it support?

Most AI clients that support remote MCP servers over HTTP with OAuth sign-in. Blackboard has tested Claude, ChatGPT and Codex, Microsoft Copilot, GitHub Copilot, Gemini, Perplexity, and Superhuman Go by Grammarly. Your institution decides which clients it permits; the server cannot tell which client or license tier is calling, so that decision is enforced through your AI vendor agreements and device and identity policies.

Turning it on and controlling access

Is it enabled automatically?

No. At general availability the MCP Server appears as a new MCP link in the Integrations section of the Blackboard LMS Administrator Panel, alongside LTI and REST, and it is off by default. Nothing connects until a Blackboard LMS administrator turns it on.

What can an administrator control?

Control

What it does

Enable MCP server

Turns MCP Server access on or off for your Blackboard LMS site.

Friendly Name (per host name)

Sets the name users see when they pick your institution during sign-in. Set in your existing host name settings.

Include in Public Host Lists (per host name)

Chooses which of your host names users can select when connecting. Set in your existing host name settings.

Can we limit which tools, or which roles, can use it?

No. The server does not offer a per-institution tool allowlist or role-based tool restriction. Access is bounded instead by each user's Blackboard LMS role and permissions: a student cannot run administrator actions, and an instructor sees only their own courses. To limit who can connect, make the Blackboard connector available only to a pilot group in your AI client's admin console, and use the client's approval policy to control which actions need a person's sign-off.

Important

Block tools in your AI client. Most institution-level AI accounts, such as business, enterprise, and education plans, let an administrator in the AI tool block specific tools for every user on that account. If your committee wants a tool unavailable, for example the bulk provisioning tools, block it there.

Can we restrict access to our enterprise AI plan only?

Not from the server. An MCP server cannot reliably identify which product or license tier a client is using. Scoping by Blackboard LMS role limits how much data a person can reach, but it does not control where that data goes. Restrict where data goes through your AI vendor, for example by permitting only institutionally managed business, enterprise, or education seats and using any connector controls your vendor offers on those plans, and through your own acceptable-use policy.

How do we turn it off?

Switch MCP off in the Integrations section of the Administrator Panel. New connections stop right away; people who are already connected keep access until their token expires.

Identity and authentication

How do users sign in?

Through the OAuth 2.0 authorization code flow. The first time a user connects, the AI client opens a browser. The user picks your institution, signs in on your normal Blackboard LMS or single sign-on pages, and approves the request. The sign-in service then issues the client a signed session token. Your existing sign-in policies, including multi-factor authentication through your identity provider, apply unchanged.

Does the AI client or Blackboard see users' passwords?

The AI client never receives a Blackboard LMS password. Sign-in happens on Blackboard and institution pages, and the client holds only a token.

Whose permissions does each call use?

The signed-in user's. The server acts as that person through the Blackboard LMS REST API, and the LMS enforces their existing permissions on every call. If a request reaches a course or role the user cannot access, the result is empty or an authorization error, never another person's data. Blackboard LMS release rules, such as adaptive release, continue to apply.

Why does turning it on require the Guest user?

Every REST API integration in Blackboard LMS is assigned a Blackboard LMS user. The MCP integration is assigned the built-in Guest user, the account with the least access, and it is set so it cannot act as any other user. This is a placeholder only. It does not allow anonymous use: every request still requires the person to sign in with their own Blackboard LMS account and runs with their permissions. The Guest user account must be available, but you do not need to allow guest access to courses.

How long do sessions last?

Access tokens expire after 60 minutes. The client can renew access without a new browser sign-in for up to 30 days, after which the user must sign in again. For current details on ending an individual user's active session, contact your Blackboard account team.

How are tenants kept apart?

Each institution is isolated. The server completes sign-in only for Blackboard LMS sites that are registered with it and whose tenant identity has been verified, which prevents a request from being pointed at an untrusted host. Data from one institution is never returned to another.

Data handling, hosting, and residency

What data passes through the server?

Only what a tool call requests and returns: for example, a list of a user's courses, a roster, or the text of a new announcement. Results go back to the AI client that asked for them. From that point on, the data is in your AI vendor's environment and governed by your agreement with that vendor.

Does the server store or cache Blackboard LMS data?

The server does not cache Blackboard LMS responses. It keeps two kinds of records:

  • Short-lived session records used during sign-in, which expire after about 10 minutes, and records that map each registered host to its tenant.

  • An audit log of every tool call, described under Auditing and monitoring. Because the log records the parameters sent with each call, it can contain personal data that a user supplied, such as a name in a create-user request.

Blackboard LMS credentials are not stored on the server. The user's Blackboard LMS access is carried inside the signed session token held by the client.

Where is it hosted?

On Amazon Web Services, in Blackboard-operated accounts. The service is deployed in multiple AWS regions across North America, Europe, Asia Pacific, the Middle East, and Africa, and requests are routed to a regional service. Ask your Blackboard account team to confirm the region that serves your Blackboard LMS site if data residency is a requirement for your review.

Is data encrypted?

Yes. All traffic to and from the server uses HTTPS. Stored records are encrypted at rest using AWS platform encryption. Session tokens are digitally signed so they cannot be altered.

Does Blackboard use the data to train AI models?

No. Blackboard does not use data passing through the MCP Server to train AI models.

AI vendors and model training

Will our AI vendor train on institutional data?

That depends on the license tier you hold, not on which vendor you choose. The server's controls stop at the boundary of the AI client. Once a result reaches an assistant, what happens to it is set by your contract with that vendor. The table summarizes what each vendor publishes; confirm the current terms with your vendor directly.

Client

Tiers that exclude your data from training by default

What to watch

Claude and Claude Code

Claude for Work (Team and Enterprise) and the Anthropic API.

Consumer Free, Pro, and Max plans follow a different policy. Submitting thumbs-up or thumbs-down feedback can opt that exchange in, even on commercial plans.

GitHub Copilot

Copilot Business and Copilot Enterprise, under GitHub's Data Protection Agreement.

Individual Free, Pro, and Pro+ plans have been trained on by default since April 24, 2026, and the opt-out is a per-user setting an institution cannot enforce centrally.

ChatGPT and Codex

ChatGPT Business, Enterprise, and Edu, and the OpenAI API platform.

Consumer ChatGPT plans follow a different policy.

Caution

Check the seats, not just the agreement. An institutional agreement covers the seats it covers. A department that expensed individual subscriptions, or a staff member using a personal account on a work machine, sits outside it. The single most important control your committee can require is that people connect only from institutionally managed accounts on a tier that excludes training.

What should we ask our AI vendor in writing?

  1. That your tier excludes your inputs and outputs from model training by default.

  2. How long data is retained, and in which regions it is stored.

  3. Which in-product actions, such as submitting feedback on a response, count as an opt-in that overrides the default.

  4. Whether the vendor offers administrator controls to limit which connectors your users can add.

Prompt injection and human oversight

What is prompt injection, and does it apply here?

Yes, it applies to any AI assistant that reads content and can take actions. An assistant does what it is told, including by content it reads. A hidden instruction in a document, an email, or a web page can steer it toward tool calls nobody intended.

What limits the damage if an assistant is manipulated?

The server bounds what any single call can do:

  • Every call is limited to the signed-in user's own Blackboard LMS permissions.

  • The catalog contains no delete tools.

  • Bulk tools are capped at 100 items per call.

  • Every call is recorded in the audit log.

These controls limit the damage. They do not stop an attempt, which is why the approval policy in your AI client matters.

Does the server require a person to confirm write actions?

No, and it cannot. Approval policy, meaning which actions need a person's sign-off before they run, lives in the AI client. It is your institution's to set and enforce, ideally centrally through the client's admin console rather than user by user.

Should a person approve every action?

The evidence says blanket manual approval is less effective than it looks. In a controlled study that Anthropic published on August 7, 2026, paid professional developers caught 13.6% of dangerous commands inserted into their work, while the classifier-based screening in Claude Code's auto mode caught 89% of the same commands. Human detection also fell over a session, from about 17% early on to about 5% after fifty or more prompts. Asked to approve everything, people approve everything.

Two caveats matter. These figures describe Claude Code, not every client: in a third-party evaluation of 720 indirect prompt-injection attempts, Claude models in auto mode stopped all of them, while Codex let 5.83% through in its auto-review mode and 19.03% in full-access mode. And Anthropic describes auto mode as a best-effort classifier that does not eliminate risk.

Sources: Anthropic, "Auto mode is now the default in Claude Code" (claude.com/blog/auto-mode-default-in-claude-code) and Anthropic Engineering, "How Claude Code auto mode works" (anthropic.com/engineering/claude-code-auto-mode).

What approval policy do you recommend?

Match the guardrail to the risk. Spend human attention where actions change data.

Type of action

Recommended setting

Read tools, such as listing courses, rosters, or assignments

Let them run under the client's automatic screening.

Write tools, such as create, update, enroll, or copy

Require explicit human approval.

Bulk tools

Require explicit human approval, and check the scope before approving.

Any client without classifier-based screening

Keep manual approval on all write and bulk tools.

Auditing and monitoring

What is logged?

Every tool call. Each audit entry records the time, the tool name, the institution, a user reference, the parameters sent, the response status, any error detail, and the call duration. The server's audit log is kept separately from Blackboard LMS logs.

How long are audit records kept?

90 days. If your institution needs to retain records for longer, request them through Blackboard Support within that window and retain them under your own records policy.

How do we get our audit records?

Open a case with Blackboard Support. There is no self-service audit view.

Is there anything the log does not show?

Yes. The server logs every call it receives, but what the assistant did with the results, and why it chose to act, happens inside the AI client and sits outside the server's audit trail. Use your AI vendor's enterprise audit and compliance features to cover that side.

Tools and write operations

What can the tools do?

Tools are organized around three roles.

Role

Examples

Student

See my courses; show my assignments and their submission status.

Instructor

Find students who have not submitted; post an announcement; open a discussion; add course content such as folders, documents, and links.

Administrator

Search users and courses; create and update courses, users, enrollments, and terms; copy a course; run bulk provisioning.

The Blackboard MCP Server Tool Guide on Blackboard Help lists every tool and marks each one that changes data.

Which tools change data?

Any tool that creates, updates, enrolls, or copies. These are the tools to point your approval policy at. Read tools retrieve information and change nothing.

Can it delete data?

No. There are no delete tools in the catalog.

Can a user call any Blackboard LMS API through it?

No. The server exposes a defined catalog of tools. It does not offer a general-purpose tool for calling arbitrary APIs.

Compliance, contracts, and cost

How does it fit with FERPA, GDPR, and similar obligations?

The server is designed to operate within FERPA, GDPR, and SOC 2 expectations. It exposes student data only to the authenticated user, within that user's existing Blackboard LMS permissions, and keeps institutions isolated. Your obligations also depend on your AI vendor agreement, because data leaves Blackboard's environment when it reaches the assistant. Your privacy office and counsel should decide how those obligations apply at your institution.

Where can we get security documentation?

Request current security and compliance documentation, such as audit reports, security questionnaires, accessibility conformance reports, and data processing terms, through your Blackboard account team. Your existing Blackboard agreements and data processing terms continue to govern your relationship with Blackboard.

Does it cost extra?

There is no additional charge to enable the MCP Server.

Approval checklist

Use this list to structure your committee's decision. Each item names who owns it.

Item

Owner

Decide which AI clients and license tiers are permitted, and confirm training exclusion, retention, and opt-in actions in writing.

Procurement and AI governance

Confirm every person who will connect uses an institutionally managed account on a permitted tier.

IT and identity

Set the AI client's approval policy centrally: reads run under automatic screening; writes and bulk tools need human approval.

AI client administrator

Decide the pilot group and make the connector available to them in your AI client.

AI client administrator

Set a friendly name and include at least one host name in the public host lists.

Blackboard LMS administrator

Review Blackboard LMS roles and permissions for the pilot group, since they bound every call.

Blackboard LMS administrator

Agree how often to request audit records, and where to retain them beyond 90 days.

Information security

Update acceptable-use guidance for faculty, staff, and students, including a ban on personal AI accounts for institutional data.

AI governance

Confirm the hosting region that serves your site, if residency matters.

Information security, with your Blackboard account team

Turn the MCP Server on in the admin panel.

Blackboard LMS administrator

Glossary

Term

Meaning

MCP (Model Context Protocol)

An open standard that lets an AI client discover and call external tools.

MCP client

The AI application that connects to the server, such as Claude, GitHub Copilot, or Codex.

Tool

A single named action the server can perform, such as "list my courses."

OAuth 2.0

The standard sign-in flow used to authorize a client without sharing a password.

Session token

The signed, time-limited credential a client uses after sign-in.

Tenant

A single institution's isolated space within the service.

Approval policy

The client setting that decides which proposed tool calls need a person's sign-off.

Auto mode

A client setting that screens each proposed tool call with a safety classifier instead of prompting a person for everyone.

Prompt injection

Instructions hidden in content an assistant reads, intended to make it take unintended actions.

Training exclusion

A vendor's commitment that your inputs and outputs are not used to train its models.