What the demonstration did
Replaying Datadog's CoPhish demonstration / published October 2025; the run date is not stated
- 1
An agent built to forward tokens
Not dated · before Oct 20, 2025The researcher registers a multi-tenant app with a client secret, points the agent's login at its consent page, adds a token-forwarding request to the sign-in topic and publishes the demo website.
- 2
Consent at the agent's login
Not dated · before Oct 20, 2025A test user clicks Login on the demo website and consents to Mail.ReadWrite, Mail.Send and Notes.ReadWrite; the post says an administrator could consent to any delegated Graph permission.
- 3
The token leaves through Copilot Studio
Not dated · before Oct 20, 2025Copilot Studio forwards the delegated token to the researcher's endpoint from Microsoft's addresses; the post shows no use of it against mail or notes.
In the staged run, the test user's consent gave the researcher's agent a delegated token with Mail.ReadWrite, Mail.Send and Notes.ReadWrite, and Copilot Studio forwarded it to an endpoint the researcher controlled; no use of the token against a mailbox or notes is shown, and no in-the-wild use is reported.
What CoPhish shows about your consent settings
The lure is a login button on a Microsoft domain, and the token leaves from Microsoft's addresses, so the link looks like Microsoft's and your users' traffic shows no connection to the endpoint. What you control is consent: with user consent turned off and the admin consent workflow on, a user cannot grant an app's delegated permissions alone, though a holder of a role that can grant consent stays outside that setting. For administrators, eligible assignments in Privileged Identity Management and a custom consent role narrow, but do not end, the exposure. If the app and the agent sit in your own tenant, turning off app registration for users, blocking password secrets and a Power Platform data policy on the HTTP and Direct Line connectors cap the build; an agent in someone else's tenant is outside all three. A token already granted stays valid until you revoke the grant.
Would the controls have stopped it?
In hindsight, no control would have stopped it. 8 controls would have limited how far it went. 2 steps have no cloud-provider control. The counts below are control assessments, not attack stages: several can address the same step. The same replay with the controls in place before the incident shows where each phase would have been refused or bounded.
- 0
- Prevented
- 8
- Contained
- 2
- Not covered
How it could have been prevented
The same replay with the controls in place. Green marks where a step is refused, blue where it is bounded, and red where no control reaches it. Open a row to read the control, why it acts, and its caveat.
Replaying Datadog's CoPhish demonstration / published October 2025; the run date is not stated
1 · An agent built to forward tokens
4 Contained1 Not covered2 · Consent at the agent's login
3 Contained3 · The token leaves through Copilot Studio
1 Contained1 Not coveredKey facts
- Datadog Security Labs published CoPhish on October 20, 2025 and updated the post on October 28, 2025 to clarify consent details; it reported no in-the-wild use
- The researcher registered a multi-tenant app with a client secret and the reply URL of the Bot Framework token service; the post said any member user can register such an app in a tenant by default
- An HTTP Request action in the agent's sign-in topic sent the user's token to the researcher's endpoint from Microsoft's IP addresses, so the user's own web traffic showed no connection to it
- The captured token carried three delegated scopes, Mail.ReadWrite, Mail.Send and Notes.ReadWrite; the post showed no use of it against a mailbox or a notebook
- As of October 28, 2025, per Datadog, users could still consent to five scopes under the default policy; an update Microsoft announced for late October 2025 would leave only Notes.ReadWrite of them, and its rollout was not confirmed
- On October 25, 2025, Microsoft told BleepingComputer it would address the technique through future product updates and that it relies on social engineering
Sources
Every fact in this breakdown comes from the public reporting below. The replay is a model of that reporting, not a reconstruction of the environment it describes.
- 1.CoPhish: Using Microsoft Copilot Studio as a wrapper for OAuth phishingDatadog Security Labs · Oct 28, 2025 · ResearchThe multi-tenant app registration with a client secret and the Bot Framework reply URL, the agent's manual authentication and authorization URL template, the sign-in topic HTTP request that forwards the token, the demo website on a Microsoft domain, the consent flow and validation code, the token's scopes, the unprivileged and administrator scenarios, the default consent policy changes of July and late October 2025, the audit log detections and the mitigations
- 2.New CoPhish attack steals OAuth tokens via Copilot Studio agentsBleepingComputer · Oct 25, 2025 · NewsMicrosoft's statement on future product updates, its framing of the technique as social engineering and its mitigation advice; no in-the-wild use reported
The estate in the replay follows the researchers' description of their own test environment. Verdicts describe what a control does when it is in place before the run. They are not an assessment of any company's security program.