Skip to content

Control 2.16: Federated Copilot Connector and Model Context Protocol (MCP) Governance

Control ID: 2.16 Pillar: Security & Protection Regulatory Reference: GLBA §501(b), SEC Reg S-P, FFIEC, OCC Bulletin 2023-17 Regulatory Applicability: Firm/counsel determination — see Regulatory Applicability Questions Last Verified: 2026-09-24 Governance Levels: Baseline / Recommended / Regulated


Scope boundary: FSI-CopilotGov vs FSI-AgentGov

This control governs the Microsoft 365 Copilot surface only — tenant-level configuration, data-source posture, audit/eDiscovery, and admin-managed extensibility. Governance of the agents themselves (Copilot Studio agents, declarative agents, Agent Builder, custom pro-code agents) — including agent registration, risk tiering, environment zoning, model-card review, and lifecycle promotion — lives in the companion FSI-AgentGov framework. See Relationship to FSI-AgentGov for the full boundary map.

Regulatory applicability

Regulatory citations identify commonly referenced requirements. Applicability, record classification, and retention periods are a firm and counsel determination; this framework does not assert that any regulation applies or is satisfied. See the Regulatory Applicability Questions appendix.

Objective

Establish governance controls for federated Copilot connectors using the Model Context Protocol (MCP) — a real-time data retrieval mechanism that is architecturally distinct from plugins and Graph connectors. Federated connectors authenticate with user credentials (not admin-managed service accounts) and retrieve live data from external sources without indexing. Microsoft-published connectors can be available by default subject to the tenant's Allowed agent types settings, while partner connectors require administrator approval. This control helps organize internal evidence for firm/counsel review against GLBA safeguard requirements, SEC Reg S-P privacy obligations, and FFIEC expectations for third-party data access governance.


Why This Matters for FSI

  • GLBA §501(b) requires safeguards for customer information systems — federated connectors introduce real-time external data flows into Copilot responses, expanding the security boundary beyond Microsoft's managed environment
  • SEC Reg S-P (248.30) requires safeguarding customer records and information — user-credential-based connectors that pull external data into Copilot responses may expose customer information to third-party services without centralized data governance review
  • FFIEC IT Examination Handbook (Information Security) expects controls over API access and external data connections — federated connectors bypass the traditional Graph connector indexing model, creating a new unindexed data flow
  • OCC Bulletin 2023-17 (Third-Party Relationships: Risk Management) rescinded and replaced OCC Bulletin 2013-29; it requires risk management throughout the third-party relationship lifecycle — each federated connector vendor (Canva, HubSpot, Notion, Linear, Google services) is a third-party data processor
  • Firms should define the primary sources, internal policies, and supervisory expectations they use to understand data sources used for AI-generated outputs — federated connectors introduce external data into Copilot responses that may not be subject to the same data quality and accuracy controls as internal data

Control Description

Cross-tenant scope: Cross-tenant federation patterns (Entra Agent ID trust, MCP federated server attestation, Copilot Studio multi-tenant publishing) are addressed by Control 2.17 — Cross-Tenant Agent Federation, which extends this control to multi-tenant scenarios.

MCP server-trust governance is authoritative in FSI-AgentGov. This control governs the Copilot-side surfacing/blocking of federated connectors through Agents > Settings > Allowed agent types and Copilot connectors > Your connections, and of MCP servers through Agents > Tools — the legitimately Copilot-managed planes. The deeper governance of MCP servers themselves (build, signed attestation, trust, and Bring-Your-Own MCP server lifecycle) is specified in FSI-AgentGov Control 2.17 — Multi-Agent Orchestration Limits and the MCP Server Governance playbook. Consume and cross-reference those controls rather than re-specifying MCP-server trust here.

What Are Federated Copilot Connectors?

Federated Copilot connectors use the Model Context Protocol (MCP) to retrieve real-time data from third-party sources. Unlike Graph connectors (which index external data into the Microsoft 365 Graph), federated connectors query external sources live at the time of the Copilot interaction.

Architectural Differences from Graph Connectors

Characteristic Graph Connectors (Control 2.13) Federated Connectors (This Control)
Data model Indexed — data is ingested into M365 Graph Live retrieval — data is fetched in real time
Authentication Admin-managed service accounts or app-only User credentials (delegated)
Data residency Data resides in M365 tenant Data resides at the external source
Admin control Admin deploys and manages Publisher-category policy plus connector-specific allowed-user scope; user-initiated authentication
DLP coverage Indexed data subject to M365 DLP policies Real-time data may not be subject to DLP until surfaced in response
Audit trail Connector ingestion events in Purview audit log Connector invocation events in Purview audit log
Default state Admin-deployed Microsoft-published connectors can be available by default; partner connectors require admin approval

Current Federated Connector Catalog

Federated connectors are generally available for Microsoft 365 Copilot per Microsoft 365 roadmap item 501120 (GA April 2026) and use MCP for real-time access with the user's identity. The roadmap states GA support for Researcher, Microsoft 365 Chat, and Agent Mode in Excel. The connector catalog is evolving, and roadmap item 570964 lists create/update/delete connector tools for Copilot Chat as In development with GA targeted for October 2026; it states those actions act with the user's account and permissions and require explicit user confirmation before running. Organizations should monitor the M365 Admin Center and Message Center for the tenant's current connector catalog, available tools, and write-action controls.

Risk Profile for FSI

User authenticates personal/work account with third-party service
                          │
                          ▼
              ┌───────────────────────┐
              │ Federated Connector   │
              │ (MCP Protocol)        │
              │                       │
              │ ├─ User credential    │  ← No admin consent gate
              │ ├─ Real-time query    │  ← Data not indexed/cached
              │ ├─ External response  │  ← May include PII, NPI
              │ └─ Injected into      │
              │    Copilot context    │  ← Becomes part of AI response
              └───────────────────────┘
                          │
                          ▼
              Copilot response includes external data
              (subject to whatever DLP policies apply
               at the response layer)

Key FSI risks:

  1. No centralized data review: Users authenticate directly with external services — there is no admin consent gate or data ingestion review before data flows into Copilot responses
  2. Personal account mixing: A trader or advisor could authenticate a personal Google Calendar or Notion workspace, introducing personal data into regulated Copilot workflows
  3. Data residency bypass: External data is fetched from the third party's infrastructure, potentially outside the firm's approved data residency boundaries
  4. Chinese wall violations: Federated connector data from one business unit's external tools could be surfaced in another unit's Copilot responses if information barriers do not extend to federated data flows

Copilot Surface Coverage

M365 Application Federated Connector Support Notes
Microsoft Copilot Chat (Researcher) Yes Primary surface for federated connectors
Microsoft Copilot Chat (Standard) Yes Roadmap item 501120 lists Microsoft 365 Chat as GA support for federated connectors
Excel (Agent Mode) Yes Roadmap item 501120 lists Agent Mode in Excel as GA support for federated connectors
Word Planned Future expansion
Teams Planned Future expansion
Outlook Planned Future expansion
Microsoft Scout (Frontier preview) — tool servers Reference only Microsoft documents Scout tool-server controls rather than MCP-specific connectivity. Local tool output is stored on the local endpoint and is outside the Microsoft 365 DPA. Include Scout's approved tool-server inventory in Control 4.16, and coordinate separate MCP approvals under this control only where the same external integration is also exposed as a federated connector.

Governance Levels

Level Requirement Rationale
Baseline Disable all federated connector access: review Agents > Settings > Allowed agent types and restrict the Microsoft-published and external-publisher categories used by federated connectors where the broader effect on agents and apps is acceptable; set every federated connector's allowed-user scope to No users; document the restriction rationale; monitor Message Center for tenant-specific action notices Maximum restriction during initial governance assessment while recognizing that publisher-category controls affect more than connectors
Recommended Selectively enable approved federated connectors; restrict availability to specific user groups via Entra security groups; require security review before enabling each connector; quarterly review of enabled connectors and usage patterns; block connectors that access personal account data Controlled enablement with formal approval — suitable for firms that need specific external data integrations
Regulated All Recommended requirements plus: full third-party risk assessment for each connector vendor; connector-specific DLP policies at the response layer; real-time monitoring of connector invocation patterns; connector usage included in examination evidence packages; annual connector security re-assessment; information barrier verification for connector data flows Comprehensive connector governance — designed for firms where external real-time data access requires formal risk management equivalent to third-party vendor onboarding

Setup & Configuration

Step 1: Assess Current Federated Connector and MCP Tool State

Portal: Microsoft 365 Admin Center > Agents > Settings > Allowed agent types; Copilot connectors > Your connections; and Agents > Tools (where licensed)

Federated connectors and MCP-based tools are governed through separate, complementary controls:

  1. Agents > Settings > Allowed agent types controls publisher categories across the tenant. For federated connectors, Microsoft documents deselecting the Microsoft-published and external-publisher categories to stop those connector categories from being enabled by default, including future connectors. This is a broad category control that also affects agents and apps in those publisher categories; assess the tenant-wide impact before changing it. Disabling the connector category automatically sets existing connectors' allowed-user scope to No users, so capture approved connector assignments before making the category change.
  2. Copilot connectors > Your connections is the connector-specific management surface. Record each connector's enabled state, allowed-user scope, and any Staged rollout group assignment. A connector can remain visible to administrators while its allowed-user scope is No users.
  3. Agents > Tools is the primary surface for AI-powered tools and Model Context Protocol (MCP) servers available to agents. This page is rolling out to Frontier tenants and may not yet be available in every region.
    • Filter by Type = MCP Server to enumerate MCP-based tools that are currently Available or Blocked.
    • Note that Microsoft documents Work IQ MCP as a single MCP endpoint with a fixed generic tool surface over resource paths, not as separate workload-named Microsoft-published MCP server SKUs. Treat any workload-specific MCP server names shown in the tenant as tenant inventory evidence, and separately track BYO MCP servers registered through the Bring Your Own (BYO) MCP server flow.
  4. Record current state in the connector and tool inventory, including publisher category, connector vendor, MCP server publisher, allowed-user scope, staged group assignments, and whether users authenticate with work or personal accounts.

Step 2: Disable or Restrict Federated Connectors and MCP Tools (Baseline)

Portal: Microsoft 365 Admin Center > Agents > Settings > Allowed agent types; Copilot connectors > Your connections; and Agents > Tools (where licensed)

  1. Review Allowed agent types. For the Baseline posture, deselect the Microsoft-published and external-publisher categories used by federated connectors where the broader effect on other agents and apps is acceptable.
  2. In Copilot connectors > Your connections, set every federated connector's allowed-user scope to No users. Selective connector exceptions begin at the Recommended level in Step 3.
  3. Block non-essential MCP servers through Agents > Tools. For BYO MCP server registration requests, enforce a "default deny" stance — reject or hold pending requests until the registration approval workflow in Step 3 is in place.
  4. Document the category-policy decision, each connector-specific access decision, and the MCP tool restriction rationale in governance records.

The tenant-wide PowerShell toggle is retired

Microsoft Learn states the Set-FederatedConnectorToggle command-line toggle for setFederatedConnectors is being retired by August 25, 2026 so connector and agent settings are governed from the same global tenant setting. The FAQ also states that if an admin previously disabled federated Copilot connectors with Set-FederatedConnectorToggle, that setting is honored until October 20, 2026, and admins must update Allowed agent types by then to continue that behavior after October 20, 2026. Do not rely on historical cmdlet output as current effective-access evidence after that transition. Use Agents > Settings > Allowed agent types, connector-specific allowed-user scope, and controlled end-user tests as the current evidence set.

Step 3: Selective Connector Enablement and MCP Server Approval (Recommended/Regulated)

Portal: Microsoft 365 Admin Center > Copilot connectors > Your connections; Agents > Tools > Requests tab

For organizations that need specific federated connectors or MCP servers, use the centralized approval workflow:

  1. For a federated connector approved through the vendor-risk process, open it under Copilot connectors > Your connections and set its allowed-user scope to the approved users or groups. Use Staged rollout where available before broader access.
  2. From Agents > Tools, select the Requests tab to review pending MCP server registration requests. Each pending request shows the server name, publisher, requested-by user, and request date.
  3. Review the server information, declared tools, and requested capabilities for accuracy and policy alignment.
  4. Select Approve to make the server available in the organizational registry, or Reject to deny the request — record the decision rationale.
  5. After Approve, the portal prompts for the Microsoft Entra permissions the MCP server requires. Review the requested permissions and grant Entra consent only if the scopes are least-privilege and aligned with the firm's approved data classifications. The server is unavailable to agent-building surfaces until consent is granted.
  6. After approval, allow up to 30 minutes for the MCP server to appear in all Microsoft Copilot Studio environments in the tenant.

Wrap the portal approval flow in the firm's broader connector and MCP server intake process:

Step Owner Deliverable
1. Business request Requesting department Business justification for specific connector or MCP server
2. Security review Information Security Data flow analysis, vendor assessment, requested Entra permission scopes
3. Privacy review Privacy/Legal Privacy impact assessment for external data access
4. Compliance review Compliance Regulatory risk assessment (GLBA §501(b), Reg S-P implications)
5. User scoping IT Operations Restrict connector/tool to approved Entra security groups
6. Configure in portal M365 Global Admin Set connector-specific access in Copilot connectors > Your connections; separately approve MCP server requests in Agents > Tools > Requests and grant Entra consent where required
7. Monitor usage Information Security Monthly review of connector and MCP server invocation patterns

Step 4: Monitor Federated Connector Activity

# Search audit logs for federated connector invocations
Search-UnifiedAuditLog -StartDate (Get-Date).AddDays(-30) -EndDate (Get-Date) `
  -RecordType "CopilotInteraction" -ResultSize 5000 |
  Where-Object { $_.AuditData -like "*FederatedConnector*" -or $_.AuditData -like "*MCP*" }

# Review connector authentication events
Search-UnifiedAuditLog -StartDate (Get-Date).AddDays(-7) -EndDate (Get-Date) `
  -Operations "UserLoggedIn" -ResultSize 5000 |
  Where-Object { $_.AuditData -like "*connector*" }

Step 5: Personal Account Policy

For FSI environments, establish a clear policy on personal account authentication:

  1. Prohibited connectors: Block connectors that primarily authenticate with personal accounts (e.g., personal Google Calendar, personal Notion)
  2. Work account only: For approved connectors, require authentication with work/enterprise accounts only
  3. User acknowledgment: Require users to acknowledge that connector-accessed data may be included in Copilot responses subject to firm retention and compliance policies

Financial Sector Considerations

  • Third-party risk management: Each federated connector vendor should be assessed under the firm's third-party risk management framework. Unlike Graph connectors where data is ingested into the M365 tenant, federated connector data remains at the third party's infrastructure — the risk assessment should address data processing, storage, and access controls at the vendor level.
  • Data residency implications: Federated connectors retrieve data from external services that may process and store data outside the firm's approved jurisdictions. Organizations should verify that each enabled connector's data residency posture aligns with regulatory and contractual requirements (see Control 2.7 — Data Residency).
  • Information barrier scope: Organizations using information barriers for Chinese wall compliance (Control 2.4) should verify whether barriers extend to federated connector data flows. A connector that retrieves data from a shared external workspace could bypass internal information barriers.
  • Audit trail completeness: Federated connector invocations should appear in the Purview unified audit log. Organizations should verify that connector events are captured with sufficient detail (connector name, data source, user identity) for examination readiness.
  • Model risk considerations: Copilot responses that incorporate federated connector data combine internal and external data sources. For model risk management purposes (Control 3.8), organizations should assess whether the provenance and accuracy of federated data meets the firm's data quality standards for AI-generated outputs.

Work IQ MCP Tenant Verification

Microsoft Learn describes Work IQ MCP as one MCP server endpoint that exposes Microsoft 365 intelligence through 10 generic tools (fetch, create_entity, update_entity, delete_entity, do_action, call_function, ask, list_agents, get_schema, and search_paths) over relative resource paths. New workloads add paths rather than new tool names. Do not document separate Microsoft-published server SKUs such as Calendar, Mail, SharePoint, OneDrive, or Teams unless the tenant inventory shows those names and the evidence is retained as tenant-specific, not public Learn fact.

FSI review should verify the actual tenant posture before enabling Work IQ in any agent scenario:

Evidence Item Why It Matters
Work IQ availability in the tenant and region Microsoft Learn notes regional allow/disallow tooling controls might not yet be available everywhere
Spending policy and billing owner Work IQ in Copilot Studio preview uses Copilot Credits / consumptive billing and requires a separate spending policy
Read/write posture Microsoft 365 Work IQ is read-only unless an administrator explicitly enables write operations
Allowed or blocked tooling controls The tenant's current Agents / Tools control surface determines whether admins can block Work IQ MCP for the scenario
Resource paths and permissions tested Generic paths may surface email, files, chats, meetings, and other Microsoft 365 resources according to user permissions and tenant policy

Preview in Copilot Studio — July 2026

Per the Work IQ in Copilot Studio (preview) documentation, adding Work IQ to Copilot Studio agents is a preview feature, uses Copilot Credits / consumptive billing, and requires a separate spending policy. Microsoft also states Microsoft 365 Work IQ is read-only unless an administrator explicitly turns on write operations in the Microsoft 365 admin center, and that allow/disallow tooling controls might not yet be available in every region. FSI organizations that enable Work IQ MCP in Copilot Studio agents should:

  • Treat preview enablement as a controlled pilot rather than production deployment, with documented risk acceptance
  • Track the preview → GA transition in the Message Center and re-review billing, DPA, support commitments, and any read-write capability before broad production use
  • Confirm regional availability of the allow/block controls in Microsoft 365 admin center > Agents and Tools; the Microsoft documentation notes this capability may not yet be available in every region
  • Preserve evidence that write operations remain disabled unless a formal risk decision explicitly enables them

Work IQ MCP server governance in Copilot Chat (surfaced through Microsoft-published tools) remains governed here; the Copilot Studio agent authoring lifecycle for Work IQ MCP servers is authoritative in FSI-AgentGov.

Defender Advanced Hunting for MCP Telemetry

Microsoft Defender XDR advanced hunting provides query-based telemetry for MCP server and federated connector activity. Organizations should configure hunting queries to monitor for:

Telemetry Signal Advanced Hunting Query Focus FSI Application
MCP server invocations CloudAppEvents where ActionType includes MCP-related operations Track which MCP servers agents are invoking and at what volume — high-volume invocations may indicate data exfiltration attempts
Connector authentication events AADSignInEventsBeta where AppDisplayName includes connector services Detect personal account authentication attempts to blocked connectors
Agent-to-connector data flow Correlate CloudAppEvents (connector invocation) with CopilotInteraction events Reconstruct the full data flow from agent request → connector invocation → response generation
Anomalous connector usage Baseline normal connector invocation patterns; alert on statistical deviation Detect compromised agents using connectors for unauthorized data access

Preview/production caveats: MCP server governance is evolving rapidly. Organizations should:

  • Monitor Microsoft 365 Message Center for changes to the Agents > Tools interface and MCP server management capabilities
  • Distinguish between GA and preview MCP servers in their governance inventory — preview servers may have different SLAs, data handling commitments, or feature limitations
  • Test new MCP server enablement in a pilot group before broad deployment

Verification Criteria

# Verification Step Expected Result Governance Level
1 Review Allowed agent types and connector-specific allowed-user scopes in M365 Admin Center Publisher-category settings match the approved tenant posture, and every federated connector is scoped to No users Baseline
2 Verify no unauthorized connector authentications Audit log shows no connector authentication events from unapproved user groups Baseline
3 Confirm connector inventory is documented Current list of connectors with publisher category, allowed-user scope, staged groups, vendor, data flow, and last review date Recommended
4 Test connector user scoping and source permissions Approved users can invoke the connector only within their source-system permissions, and unapproved users cannot connect or invoke it Recommended
5 Verify third-party risk assessment exists Each enabled connector vendor has a current risk assessment on file Recommended
6 Test DLP policy coverage for connector data Connector-sourced data in Copilot responses triggers applicable DLP policies Regulated
7 Verify information barrier scope Connector data flows respect information barrier boundaries Regulated
8 Review connector usage audit trail Connector invocation events appear in Purview audit log with required detail Regulated
9 Confirm examination-ready documentation Connector inventory, risk assessments, and usage reports are packaged for examination Regulated
10 Run quarterly connector re-assessment All enabled connectors have been re-assessed within the last quarter Regulated

Additional Resources


FSI Copilot Governance Framework v1.8.0 - July 2026