
The M365 outage took Defender XDR and Purview down with the mail they protect. The harder question is who declares failover, and at what hour.
The Microsoft 365 outage cost most Australian organisations a working day of email. It cost security teams something nobody has counted: a working day without detection.
Every conversation I have had since has been about access. When did mail come back. What did we tell staff. Do we need a second platform. All fair questions, but there's a more important one.
The outage began around 18:00 UTC on 12 November, about 5am AEDT on the 13th. Microsoft did not report restoration across all remaining regions until early on the 14th, past 27 hours. Its early fixes did not hold, and an attempt to add capacity stalled because machines stuck in a bad state blocked the changes from syncing (The Register).
Microsoft called it a technical issue unrelated to cyber or AI. A problem from an earlier network capacity upgrade surfaced after a recent network change. Some servers failed, and the remaining servers were overwhelmed by the redirected traffic (Forbes).
Exchange Online was the shared dependency. Outlook, Teams, SharePoint, Copilot and Defender XDR degraded with it under advisory MO1550821 (Cybersecurity News). In Australia the Australian Federal Police, ASIO and the ABC were among those affected (Information Age).
Read that service list again. Defender XDR is on it.
Cloud Security Alliance Labs notes that Microsoft's advisory later extended to Defender for Office 365, including Safe Links, with Purview data loss prevention and auto-labelling also reported as affected. Its argument is that these were not independent product failures. They shared an upstream dependency (CSA Labs).
So for most of a working day, the controls that watch the suite ran on the suite. Detection, link protection and data loss prevention degraded alongside the mail they protect. A lot of organisations had no reliable view over their primary attack surface, during a well-publicised event that every phishing operator in the world was watching.
I have not seen anyone report that as the incident.
Two jobs follow from it, and both are still open at most places I have spoken to. Hunt the telemetry gap now that detection is back. Then check what your queued automation and agents did with a day of delayed mail.
Teams voicemail lives in Exchange Online mailboxes. The admin centre you would use to reroute was affected. Service health notices arrive through the tenant that is down.
A continuity plan built on product names models none of that. It will look complete right up to the moment you test it, because the failure does not travel along product lines. It travels along shared dependencies: mail transport, Graph, identity.
Rewrite the plan around those three and the questions change. Think about what still works when identity is degraded.
Microsoft's first fixes did not work. Added capacity could not sync. The restore ran gradually, which made it slower.
That pattern is normal in a correlated failure, and it is missing from most recovery time objectives. An RTO of four hours usually assumes recovery runs back through the failure at roughly the same speed. It does not. Systems fail quickly and recover slowly, and a provider restoring carefully will take longer than one restoring recklessly.
So a 27-hour outage against a four-hour RTO is not so much a bad estimate, it's cause to rethink your definition of RTO.
Here is the question I could not answer well on Friday, and neither could anyone I asked.
Who declares failover, and at what hour?
Microsoft kept saying relief was coming and gave no completion time. At hour three, failing over feels premature and you would look foolish if it came back at hour four. At hour twenty it feels pointless, because surely it is nearly over. So the hours pass and nothing happens, which is a decision made by default.
No product fixes that. A written trigger can, and three lines will cover it:
Now's the time to write it - before the next outage. The value here is not the threshold itself, because you will probably argue about the number for a week and probably end up setting it too high. The value is the discussion that happens now, with time to think, instead of at hour six with everyone talking over each other at once.
If you write a trigger, you need somewhere to fail over to. This is where I would be careful with the vendor claims.
A standby is not a backup. Backups do not help during an outage, because your data was never lost. Your access was. A standby is a provisioned, pre-synced place for critical people to keep working, deployed before the event rather than after it. Google sells this as Google Workspace for Business Continuity, with a Business Continuity Plus tier that syncs mail, calendar and chat (Google).
The model is sound. Three things are worth knowing before you sign anything:
The switch is manual, so pre-stage what you can. Google's guidance is to leave MX pointing at the primary until you are actually in an outage (Google), which means someone changes DNS during the incident. The change itself is quick. What fails is the preparation around it. Keep the MX TTL low rather than wherever the last migration left it, publish SPF and DKIM for the standby in advance so your first failover message does not fail DMARC, and make sure more than one person can get into the registrar at 3am.
You get recent history, not your archive. The usual sync seeds with the most recent 30 days. That is fine for continuity. It is not fine if nobody warned your executives.
There are caps. Business Continuity Plus allows 30 consecutive and 60 cumulative days a year, with up to four failover tests annually (Google).
CPS 230 has been in force since 1 July 2025. It requires entities to maintain a register of material service providers and to manage the risks of relying on them (APRA).
One suite now carries productivity, collaboration, AI assistance and security controls. That makes it a single material service provider sitting underneath several critical operations at once. Last week is the evidence for a register entry that a lot of organisations wrote optimistically, and your risk function will have a view on how that entry should read now.
Disclosure first: Restack is partnering with a platform called CloudM in this space.
CloudM Continuity automates the replication, failover and failback between Microsoft 365 and a Google Workspace standby, on top of the Business Continuity Plus licensing above (CloudM). It is the most complete version of this model I have seen. The reverse sync that reconciles work done during an outage solves a real problem, because do-it-yourself failover tends to leave you with two halves of a mailbox and no way to merge them.
It is also new, and mail is the first data type available. I would rather say that plainly than oversell it. We are running it ourselves before we put a customer on it. Feel free to have a chat if you're interested.
Of course, this does not remove your dependency on Microsoft and it is not meant to. Leaving Microsoft altogether is a different project with a much larger invoice, and I can't see many orgs heading down this path.
Four things, none of which need a purchase order.
Then test it. Four tests a year is a cap on the licence. It is not a target.
Microsoft had not published a full post-incident review for MO1550821 when this was written, so the cause above is its own interim statement.
Tell us what's not working in your environment. We'll tell you what we'd do about it, or point you in the right direction.