AI security & operations for MSPs

Entra Sign-in Paths

Hal reads every Microsoft tenant’s Entra sign-ins from Microsoft Graph. An Event Hub can also be set up in the client’s own Azure, to which Microsoft delivers the tenant’s sign-ins as they happen. Hal reads the hub beside Graph, while the tenant’s sign-ins keep coming through Graph until you switch the tenant to the hub. Everything else Hal reads from Entra (directory audits, security alerts and risk detections) comes through Microsoft Graph either way. What is collected is on Entra ID.

Ask Hal in the portal chat about a tenant’s sign-ins. He tells you how they are arriving now and, for an Event Hub, what it needs and Azure’s current price, and you decide whether to set one up, and later whether to switch to it. Hal never sets up a hub, or switches a tenant to one, on his own judgement.

Microsoft Graph

Graph needs nothing beyond the connection you already made for Microsoft 365. Hal polls each tenant every few minutes.

Microsoft paces how fast an app may read a tenant’s sign-ins, and on a busy tenant it slows or refuses Hal’s requests, so sign-ins can arrive late, by hours when the pacing is heavy. None is lost: Hal reads on from where he stopped, as far back as the 29 days Microsoft serves. Hal can tell you, for each tenant, how far behind the clock his reading of its sign-ins is and how often Microsoft refused his requests over the last day.

An Event Hub in the client’s Azure

Through a diagnostic setting in Microsoft Entra, Microsoft can deliver a tenant’s sign-ins to an Event Hub. Hal reads the hub beside Graph. Microsoft requires the hub to belong to the tenant whose sign-ins it carries, so it sits in the client’s own Azure subscription.

What it needs

  • An Azure subscription in the client’s own Microsoft tenant. Without one, a tenant cannot have a hub.
  • One sign-in, with a device code, by someone holding Owner on that subscription (or Contributor with User Access Administrator) and the Security Administrator role in Microsoft Entra.
  • A monthly Azure charge to the client. The hub runs on Event Hubs Standard, starting at one throughput unit and raised automatically to at most four when the sign-ins need it, billed by the hour at the units in use, plus Microsoft’s charge per million events. Azure never lowers the units on its own; ask Hal and he tells you how many the hub holds and what they cost, so you can lower them in the client’s Azure. Azure’s Event Hubs pricing has the rates; ask Hal and he gives you the current list price, read from Azure’s price list at the moment you ask. The client’s own agreement with Microsoft may differ.
  • Time for the first delivery. Microsoft takes 24 hours to 3 days to deliver a new hub’s first sign-ins.

Setting it up

A Hal admin asks in the portal chat, for example:

“Send Acme’s Entra sign-ins to an Event Hub”

Hal shows a device code, and the person with the roles above signs in with it. Hal lists the subscriptions that sign-in sees in the client’s tenant, checks each one, states the monthly charge, and creates nothing until you confirm a subscription. Then he creates, in this order:

  1. the resource group hal-entra-signins, in East US;
  2. the Azure Event Hubs Data Receiver role on that resource group, for the app registration Hal already reads the tenant with;
  3. the Reader role on that resource group, for the same app registration, so Hal can see the namespace’s throughput units, its incoming sign-ins and any throttling;
  4. the resource providers Microsoft.EventHub and Microsoft.Insights, registered in the subscription, each only where the subscription lacks it;
  5. an Event Hubs namespace, Standard tier, starting at one throughput unit and raised automatically to at most four, which Azure never lowers on its own;
  6. an event hub, entra-signins, which keeps 7 days;
  7. a consumer group, hal, and an access rule on the namespace, which the diagnostic setting uses;
  8. Microsoft Entra’s diagnostic setting hal-entra-signins, which sends the tenant’s four sign-in logs (interactive, non-interactive, service principal and managed identity) to the hub.

The sign-in’s write access serves the setup and is not kept. Afterwards Hal keeps only two read-only roles in the client’s Azure: Azure Event Hubs Data Receiver, to read the hub, and Reader, to see the namespace’s throughput units, its incoming sign-ins and any throttling. Once a tenant’s sign-ins come from its hub, a minute in which Microsoft throttled the hub is read from Graph.

If Azure refuses a step, for example for a missing role, the setup stops there and Hal makes no further change. He names what he made. Fix what was refused and ask him again, which finishes the setup in the same subscription, or delete what he made yourself.

Where the sign-ins come from

Setting the hub up does not change where the tenant’s sign-ins come from: they keep arriving through Microsoft Graph. Once Microsoft delivers, Hal reads the hub beside Graph, from the beginning of what it holds, and compares what he finds there with what Graph delivered. The tenant stays on Graph until you switch it, below.

What the hub’s sign-in records carry

The hub carries Microsoft’s own record of each sign-in, not Graph’s copy of it, and the two differ in a few places:

  • No MFA detail. On the rare sign-in with a fresh MFA prompt, Graph’s record names the prompt’s method. Microsoft has deprecated that detail and leaves it out of the hub’s records, so Hal reads its absence as telling nothing either way. Each authentication step and its method are in both records.
  • No failure reason on a successful sign-in, where Graph’s record says “Other.”. A failed sign-in carries its reason in both.
  • The exact time, to a fraction of a second, where Graph’s records give whole seconds.
  • On an application’s sign-in, the network and device values Microsoft records, which Graph’s records leave empty.
  • On a sign-in the tenant has not enabled for Identity Protection, a risk level of “hidden” where Graph’s record can say “none”. Neither is a risk.

The switch

Switching a tenant’s sign-ins to its Event Hub is your decision, made in the portal chat. A Hal admin asks, for example:

“Switch Acme’s Entra sign-ins to the Event Hub”

Hal gives you the facts first: how long the hub has delivered; how many sign-ins Microsoft delivered to it over the last day, how many of them more than an hour after they happened, and the longest delay; for each kind of sign-in, how its records in the hub match the ones Graph delivered; whether the hub’s records carry risk scores; and the tenant’s sign-ins from its trusted locations over the last 7 days.

What refuses a switch

Hal switches only when each of these holds, and says, word for word, which do not:

  • Every part of the hub has been read from its beginning to its end. The switch is placed by the oldest sign-ins the hub still holds, and every kind of sign-in it holds must be known.
  • Every kind of sign-in matches Graph’s. For each kind the hub delivered, Hal has matched each sign-in’s id with the ones Graph delivered, every sign-in Graph has had a full day with, and at most one in a thousand found no match. A kind that matched is one Hal can read from either path and store once.
  • Every sign-in in the hub carries an id. Graph’s copy of every sign-in carries its id, so Hal cannot match a hub record that has none, and storing it from the hub could store it twice.
  • Trusted locations show. If the tenant had sign-ins from its trusted locations over the last 7 days, the hub’s records have shown a trusted location, so Hal can mark those sign-ins trusted as he does Graph’s.

Each of these guards against a sign-in lost, stored twice, or stored without a fact Graph’s record carries, so none is yours to waive. The hub must also be delivering: one Microsoft has not yet delivered to cannot be switched.

How quickly Microsoft delivers is not among them. Microsoft publishes no maximum delay for the hub, and a sign-in that arrives late is examined when it arrives. Hal reports the delays; whether they suit the client is yours to weigh.

When you confirm

Hal checks every condition again at that moment, then records the switch time: the newest of the hub’s parts’ oldest events, or seven days back when that is later, plus an hour. Sign-ins created from that time come from the hub, and Graph stops reading them. The hub also sends any earlier sign-in of its seven days that Hal never stored from Graph, when it is of a kind that matched. Hal starts reading the hub within about a minute, from the beginning of what it holds; a sign-in both paths delivered is stored once. Should a sign-in reach the hub later with no id, Hal does not take it from the hub: he reads its span from Graph, by id. A kind of sign-in the hub first delivers after the switch is read from Graph the same way until Hal has checked the first twenty of its sign-ins and found each in Graph under the id the hub gave it; from then on it comes from the hub. If one is not found, that kind stays on Graph until the tenant returns to Graph. Should the hub no longer hold sign-ins from after the switch time by then, Hal reads that span from Graph, so none is lost.

If the switch cannot be written in full, the tenant’s sign-ins stay on Graph and Hal tells you; ask him again later. He does not retry on his own.

Going back to Graph alone

A Hal admin asks in the portal chat to return a tenant to Microsoft Graph alone, and confirms. No sign-in is needed. Hal stops reading the hub, and the tenant’s sign-ins go on arriving through Graph.

For a tenant whose sign-ins had come from the hub, Hal reads them from Graph again from the earlier of two moments: the last before which he had read every sign-in Microsoft had put in the hub, and the arrival of the newest sign-in the hub delivered. Graph also re-reads the day before that moment for sign-ins Microsoft delivered late, and each span of two hours or more in which the hub delivered none of one kind of sign-in, unless that span ended before the return with Microsoft’s own sign-in log showing none of that kind throughout it. A sign-in both delivered is stored once. Four cases are read by neither: a sign-in Microsoft delivered to the hub more than a day after it happened that Hal had not yet read from the hub at the return, which is very rare; a span shorter than two hours in which Microsoft dropped the tenant’s sign-ins, or one kind of them, rather than delivering them late, as far as that span lies earlier than a day before that moment; a sign-in of one kind still on its way to the hub when Microsoft stopped delivering that kind, where Microsoft’s own sign-in log then showed none of that kind until the hub delivered it again, which is very rare; and a sign-in Microsoft published to Graph more than a day after it happened, which Graph’s own reading of any tenant misses too.

Hal changes nothing in the client’s Azure, and never deletes anything there. Everything he made stays: the resource group with the namespace in it, his roles on it, and the diagnostic setting, which goes on sending the tenant’s sign-ins to the hub while it exists. Azure bills the client for the namespace until it is deleted. When the client wants it gone, they delete it themselves, by the names Hal gives: the resource group (it holds the namespace, with its event hub, consumer group and access rule), Hal’s roles on it, and the diagnostic setting, which is Entra’s and not in the resource group. The provider registrations cost nothing, other resources may use them, and they can stay. Asking Hal for the Event Hub again reuses what is there.

Related: Entra ID and Microsoft 365.

See what Hal surfaces on your own clients.

No deck. Ask us anything first. When you want to see Hal on your own tenants, we sign a short evaluation agreement and stand up your instance; you connect one tenant, Hal watches it for 14 days, and we walk you through what he found.

  1. 01You ask us your questions. No deck, no demo dataset.
  2. 02You sign a short evaluation agreement, and we stand up your own instance.
  3. 03You connect one tenant from your own admin console. Hal watches it for 14 days.
  4. 04We walk through what he found. Keep going month to month, or revoke the scopes yourself and stop.