Guides/Playbook

Copilot CLI adds Entra sign-in: Verify inference and local data

7 min read
On this page
Two separate conversation paths connect to one shared local document

Copilot Entra sign-in

General Analysis

An employee signs in to Copilot CLI with a company Microsoft account. That alone does not tell the security team where model requests go or which policies govern them.

GitHub's November 17, 2026 Entra authentication preview adds two routes: silent sign-in to a linked GitHub Enterprise Managed User, or access to Microsoft 365 Copilot inference. The preview requires CLI 1.0.113 or later on Windows or macOS; Linux support is still forthcoming. Before enabling it, name the approved inference service and decide how the local working directory and connected tools will be governed.

The setup details come from GitHub and Microsoft documentation. The pilot checks are recommendations; General Analysis has not tested this preview in a customer tenant.

Choose the inference route before approving the rollout#

The release announcement distinguishes these entitlements:

Intended routeEligibilityInference and usage authority
Entra → GitHub managed userOIDC-backed Enterprise Managed Users, assigned Copilot Business or Enterprise license, Copilot CLI policy enabledGitHub inference and the organization's GitHub Copilot plan
Entra → Microsoft 365 CopilotFrontier participation, Microsoft 365 Copilot license, usage-based billing and required admin settingsMicrosoft 365 inference, model access and usage controls; no GitHub account required

For a user entitled to both, approve the routes individually. Record the tenant, GitHub enterprise if applicable, intended service, allowed work, and administrator responsible for disabling access. A company email address is not enough information for that record.

Keep ordinary coding-agent controls around either route. Changing who supplies inference does not answer whether the agent may publish a repository, read a confidential export, or send a message.

See how your AI systems hold up under real attacks

General Analysis maps AI applications and agents, red teams prompts, retrieval, tools, MCP servers, browser actions, permissions, and business workflows, then turns findings into evidence your team can reproduce and retest.

Configure the relevant administrators' controls#

For the GitHub route, Enterprise-Managed Authorization (EMA) must be enabled in both GitHub and Entra. In the GitHub enterprise, open Identity provider → Enterprise-Managed Authorization, then review the default and overrides. A directly targeted user override wins. Without one, any applicable disabling override denies access; if all applicable overrides enable EMA, access is allowed. The default applies only when no override applies.

In Entra Enterprise apps, verify the Copilot CLI application ID 5ecc71bb-0fa6-4d6b-a31b-de12184c4b1a. Its Security → Permissions page should show admin-consented user_impersonation access to GitHub Enterprise Managed User (OIDC). Use the official setup flow to grant that relationship. Its permission entry also provides a revocation action.

For the Microsoft 365 route, GitHub's authentication guide names three settings:

  • Microsoft Copilot Code spending policy: Billing Administrator or Global Administrator.
  • Copilot Frontier: AI Administrator or Global Administrator.
  • GitHub Copilot access (Frontier): Global Administrator; disabled by default.

Review the spending policy's population and covered services. Microsoft documents an Auto-apply new services setting that is enabled by default. Turn it off for a pilot policy if future services need a separate review. Otherwise a policy approved for today's workload can acquire a broader service scope later.

After setup, run /login inside Copilot CLI and select Your organization (Microsoft Entra). Inspect /model for the available entitlements. Also review the launcher's provider configuration: GitHub documents that BYOK provider variables take precedence for model requests regardless of GitHub authentication. Do not dump environment variables or tokens into the acceptance record.

A new inference session can still read an old file#

GitHub's Microsoft 365 model documentation establishes three boundaries:

  1. A session uses one inference service at a time. Switching from Microsoft 365 to GitHub starts a new session without carrying the existing Microsoft 365 conversation context into GitHub inference.
  2. Commands and file operations run locally. Session separation does not isolate the device or its tools.
  3. Local conversation history and generated files fall outside Microsoft 365 service-side retention, legal hold, and eDiscovery policies.

Consider an illustrative workflow: an employee uses Microsoft 365 context to draft a confidential account summary and saves it in a project directory. Later, they start a GitHub inference session in that directory. The conversation reset does not remove the saved summary. If the later session is allowed to read it, that file provides a separate path into the later task.

Set a working-directory rule before the pilot. A task that handles restricted work data should have an approved storage location, retention owner, and outbound-sharing policy. If the organization requires separation between inference services, enforce that separation through workspaces, device controls, and tool permissions. Asking employees to start a new chat does not enforce the file boundary.

Use the managed-permissions guide to verify applicable policy delivery and operation decisions. For this preview, record which controls actually reach each account and session instead of assuming that a GitHub-licensed account and a Microsoft 365-only account receive identical server policy.

Review Work IQ as access to actions as well as context#

Microsoft's Work IQ MCP overview includes tools for reading, creating, updating, and deleting Microsoft 365 entities. A connection introduced as work context can therefore expose more than search.

The permission reference distinguishes WorkIQ.Read, WorkIQ.Act, and WorkIQAgent.Ask. All listed permissions are delegated and require admin consent. WorkIQAgent.Ask includes read and write access through Work IQ agents, scoped to the signed-in user. Its name is not a read-only guarantee.

Inspect the actual connection's granted scopes and exposed actions. For a pilot intended only to retrieve information, identify how writes are prevented before approving it. A user having permission to send mail does not mean every agent task should inherit permission to send it without review. Keep the tool's resource checks and the agent's approval requirements in the MCP security review.

Verify each route on a pilot device#

Use a managed pilot device and synthetic documents. These are proposed acceptance checks, not instructions to probe another tenant:

  • Start an approved route and record the selected service, model, identity, and effective policy. Compare a second route only if it is separately authorized.
  • With harmless context, check the documented session transition. Inspect session state; a model failing to recall a phrase is weak evidence of isolation.
  • Save a dummy file in the pilot workspace and check its accessibility from the later session under the intended file policy. If the requirement is separation, require an enforced denial or a separate workspace.
  • For Work IQ, confirm an allowed read and the expected restriction on an unapproved action using a disposable fixture. Preserve the tool decision and downstream result.

The following is an illustrative record format, not a Copilot configuration file or vendor response. Null fields deliberately mean evidence has not been collected. Use one record per route and attach references to redacted evidence rather than prompts, tokens, or business documents.

Code source: illustrative.

JSON
{ "device_group": "managed-pilot", "client_version": null, "tenant_id": null, "inference_service": null, "selected_model": null, "policy_evidence": [], "workspace_boundary_result": null, "work_iq_granted_scopes": [], "work_iq_action_result": null, "local_retention_owner": null, "disable_access_result": null }

Finally, have the responsible administrator disable the pilot's intended access route and verify a fresh request is denied. Check remaining credentials and entitlements separately. Stop running work and review its effects before closing the exercise; removing a sign-in permission is not evidence that a local process has stopped.

Attach the completed record to the rollout approval. If a route still permits an action the policy was meant to prevent, keep that route in the pilot while its owner corrects the restriction.

Frequently asked questions

  • Does Entra sign-in mean Copilot CLI uses Microsoft 365 inference?

    No. Entra can authenticate a linked GitHub Enterprise Managed User, which uses GitHub inference, or enable eligible Microsoft 365 Copilot models. Check the service selected for the session and the corresponding administrative controls.

  • Does switching from Microsoft 365 to GitHub inference isolate local files?

    No. GitHub documents a new session that does not carry the existing Microsoft 365 conversation context into GitHub inference. Local files, commands, and connected tools have separate access controls. A file saved during the earlier session can remain available to later work.

  • Is WorkIQAgent.Ask a read-only permission?

    No. Microsoft's permission reference says WorkIQAgent.Ask includes read and write access through Work IQ agents, scoped to the signed-in user. Review the granted scopes and available actions instead of assuming that an Ask permission only retrieves information.

Browse all