Blog

Microsoft 365 outage: thirty hours, one network change and a phishing wave

On Thursday, 12 November, Exchange Online stopped working worldwide, and with it Teams scheduling, SharePoint, Copilot, parts of Defender and the admins' own tools. Microsoft's incident record gives an impact window of 30 hours. The cause, as far as Microsoft has explained it, was not an attack but a latent bug from a network upgrade. Attackers still made use of it: by Friday they were sending phishing mails that quoted the real incident number. Switzerland mostly lost a Friday morning, because the outage began after office hours here. The next one may not.

What happened

According to Microsoft's Service Health record EX1550543, the impact began at 10:05 UTC on 12 November. Microsoft only posted about it at 16:20 UTC, 17:20 in Switzerland, when users could no longer reach their mailboxes through several connection methods. Two hours later Microsoft opened a second incident, MO1550821, for everything that depends on Exchange: Teams, SharePoint, Microsoft 365 Copilot and Defender XDR. According to The Register, Intune and the Microsoft 365 Apps admin center were down as well, so admins lost some of the tools they would have used to investigate.

The first night shows engineers working without a diagnosis. They redistributed traffic, paused all deployments, bypassed health checks and rebuilt servers with an older configuration, and restored servers failed again as soon as they took traffic. What held was a patch to an upstream authentication and routing component at 07:25 UTC on Friday, rolled out first to the regions where the working day was beginning. The record closes the impact window at 16:10 UTC on 13 November; the last regions were restored early on Saturday.

12 to 14 November 2026, UTC

Thirty hours without reliable Exchange Online

  1. 12 Nov, 10:05Impact beginsAccording to the header of incident EX1550543, set after the fact. Nobody outside Microsoft knew yet.
  2. 16:20First admin center postUsers may be unable to access their mailboxes. In Switzerland it is 17:20, end of the working day.
  3. ~18:00Outlook reports peakAbout 130,000 on Downdetector, mostly from the US, from 74 countries within an hour.
  4. 18:26MO1550821 for the whole suiteTeams, SharePoint, Copilot and Defender XDR depend on Exchange and fail with it.
  5. 19:06All changes pausedTraffic is redistributed and health checks bypassed. Restored servers fail again under load.
  6. 13 Nov, 01:06A network change is revertedFive hours after Microsoft had said no recent change correlated with the start.
  7. 07:25The patch that holdsTo an upstream authentication and routing component, rolled out first to regions entering their working day.
  8. 12:00Europe improves13:00 in Switzerland. Mail queued for up to 48 hours starts to arrive.
  9. ~15:00Chlorine accident at PhantasialandThe warning mails of a faulty dosing system never arrive. 15 hotel guests are examined, four go to hospital.
  10. 16:10End of the impact window30 hours and 5 minutes after the start in the record.
  11. 14 Nov, 06:09Restored for everyoneMicrosoft's last post on X, for the remaining customers and regions.
For Swiss time, add one hour.

Published durations range from 18 to 36 hours, depending on which start and end a source uses. 30 hours is the window in Microsoft's own record; most users were affected for about a day.

What Microsoft says caused it

Microsoft's explanation changed several times during the incident. First it reported elevated load, then servers that were unexpectedly unavailable, then servers failing automated health checks, then an overwhelmed core process. At one point it said that no recent change correlated with the start; five hours later it reverted a network configuration change. On Friday evening came the sentence that has held so far: "The incident was introduced by an issue during an earlier network capacity upgrade that surfaced following a recent network change." The final entry on Saturday speaks of a latent bug in the systems that establish secure connections to the service, exposed by traffic shifts during standard network maintenance.

EX1550543, preliminary

How a network change took down the whole suite

  1. A capacity upgrade leaves a latent bugIn Microsoft's words, an issue during an earlier network capacity upgrade. Nothing happens yet.
  2. A routine network change shifts trafficStandard maintenance moves traffic and exposes the bug in the systems that establish secure connections to the service.
  3. The connection layer is overwhelmedRequests queue up until the servers run out of processing capacity.
  4. Health checks take servers outServers fail automated health checks and leave rotation. The rest get their load and fail too.Remedy from SRE practice: Load shedding, and a cap on how much capacity failing health checks may remove.
  5. Everything built on Exchange fails with itTeams scheduling, SharePoint, Copilot, Defender, Purview and Intune, so admins also lose their own tools.
  6. Restored servers collapse againThe backlog hits them as soon as they take traffic, until a patch to an upstream component holds on Friday morning.Remedy from SRE practice: Ramp traffic up gradually after the load has fallen.
Our reading of Microsoft's status updates; Microsoft has not published this chain itself.

So the failed health checks were a symptom, not the cause, even though Microsoft first suspected them. Microsoft says there was no indication of malicious activity and that the outage was not related to cyber or to AI and agents. It promised a preliminary post-incident review within two business days and a final one within five, so around 17 to 20 November. The Azure status history lists no incident for those days. Microsoft has not named the component it patched, so whether identity or routing services outside Exchange were involved remains open until the review is out. Australia's home affairs minister compared the outage to CrowdStrike and spoke of a faulty update; Microsoft's own wording points to a network change, not a software update.

Who was hit

Within about an hour, Downdetector logged reports from 74 countries and territories. Outlook peaked at about 130,000 reports around 18:00 UTC, most of them from the US. The documented damage fell where people depend on email to work: according to the compilation on Wikipedia and the sources it cites, NHS hospitals in the UK cancelled some operations and went back to paper, a criminal trial in Manchester was suspended, offices of the US House of Representatives lost a day of legislative work, courts in New York and Maine could not rely on emailed filings, New Zealand paused visa processing, and bank staff in Hong Kong were without email. Airlines lost email and some ticketing functions, but flights ran normally.

Switzerland was shielded by timing, not by architecture. The outage started on Thursday evening Swiss time, and Microsoft reported Europe improving from 13:00 on Friday. 20 Minuten counted more than 100 reports on allestörungen.ch at 06:00 on Friday. We found no Swiss authority or company that reported concrete damage, and no statement from the Federal Office for Cyber Security. In Germany, Borns IT-Blog collected reader reports of a disrupted Friday morning; administrations that run Exchange on their own servers were not affected.

The most direct consequence close to us happened in Brühl near Cologne. On Friday, 13 November, a faulty dosing system pumped too much chlorine into the pool of the Hotel Ling Bao at Phantasialand. The system raises an alarm when the values are out of range, but only by email, and according to Phantasialand those mails never arrived because of the Microsoft outage. When the fire brigade arrived at around 16:00, guests already had irritated airways. Paramedics examined 15 people and four went to hospital, one of them moderately injured. The defect caused the accident; the outage made sure nobody heard about it in time.

Attackers were faster than the fix

For us the most important consequence is not the outage itself. According to Microsoft Threat Intelligence, as reported by Cybersecurity News, a credential phishing campaign started on 13 November that imitates Microsoft 365 service health notices and quotes the genuine incident number EX1550543. The mails claim that held messages will be deleted within 24 hours unless the recipient restores them. Subjects follow the pattern "Outage – " plus a six-character code, the sender names are "Outage Notification" or "Service Health", and sender domains include moments-notice[.]com, planetmaison[.]com and newbooker[.]com. The reported targets were US organisations, and whether credentials were actually stolen is not confirmed.

A lure with a real incident number is credible to exactly the people who spent the night refreshing Service Health: admins. At the same time, mail filtering, link checking and parts of Defender were impaired, because they run on the same platform. A Microsoft 365 outage is therefore not only an availability problem but a security incident scenario: protection and visibility fail at the moment the phishing starts.

Not a one-off

Every large outage since 2024 for which a root cause has been published started with a change the provider made itself, and none was an attack. In most of them, health checks, retries or the recovery backlog multiplied the original fault. Google's SRE book describes this loop, and AWS reacted to its own outage in October 2025 by capping how much capacity failing health checks may remove. Microsoft 365 alone has now had three such cascades in two years.

Large outages since 2024

Each one started with the provider's own change

  1. 19 Jul 2024CrowdStrikeA faulty content update for the Falcon sensor. More than eight million Windows systems in boot loops, fixed by hand.
  2. 25 Nov 2024Microsoft 365A change linked to decommissioning a backend service, then a flood of retries. One to two days.
  3. 20 Oct 2025AWS us-east-1A latent race condition in DNS automation. About 14.5 hours, made worse by health checks that removed capacity.
  4. 29 Oct 2025Azure Front DoorAn invalid configuration got through validation. About 8 hours.
  5. 18 Nov 2025CloudflareA database permission change doubled a generated file. About 6 hours of errors worldwide.
  6. 22 Jan 2026Microsoft 365Reduced capacity during maintenance in North America, and rebalancing created new imbalances. About 10 hours.
  7. 12 Nov 2026Microsoft 365A latent bug from a capacity upgrade, exposed by a network change. About 30 hours.
None of these outages was a cyberattack.

A day without Microsoft 365 should therefore be planned for as something that happens about once a year, not as a black swan. The Cloud Security Alliance makes the point that matters most for planning: products bought separately failed as one. Copilot lost its data, Defender consoles were impaired and Safe Links could not check URLs, because they all depend on Exchange Online, Graph and identity upstream. Concentration risk has to be mapped by shared upstream service, not by product name.

What the SLA pays

Microsoft promises 99.9 percent monthly availability in its online services SLA. Credits are not automatic: they must be claimed, for licences bought through a partner via the partner, and by the end of the month after the incident, so by 31 December 2026. They only cover the affected service. Whether the outage counts as 24 or 30 hours, a fully affected user had about 96 percent availability in November, which by our reckoning means a credit of half a month's fee for Exchange Online. That is a few francs per user against a lost working day. Forrester draws the same conclusion: the SLA does not insure the business.

For banks and other institutions supervised by FINMA this outage is a ready-made test case. FINMA Circular 2023/1 has required them since 1 January 2026 to define an interruption tolerance for each critical function. If client communication, trade confirmations or incident coordination run on Exchange, Teams or SharePoint, they have to show that they stay within that tolerance through a day or two without all of them at once. In the EU, DORA asks the same of financial firms, and that Microsoft is directly overseen as a critical ICT provider does not take that responsibility off them.

What to do now

  1. Look for the phishing mails. Search for subjects beginning with "Outage –", the sender names "Outage Notification" and "Service Health", mails citing EX1550543 and the domains above. For anyone who clicked, check sign-ins, new inbox rules and OAuth consents from 12 November on.
    In CyberTell: employees report suspicious mails under "My tickets", Phishing analysis examines the headers, links and attachments of an .eml file, and Dark web monitoring looks for leaked credentials. With your own template in the style of a Service Health notice, you can later practise the same lure in Phishing campaigns.
  2. Check what the delayed mail triggered. Microsoft held incoming mail for up to 48 hours instead of bouncing it, so it arrived in a burst. Look at Power Automate flows, agents and other automations for duplicate actions, and at deadlines that ran out while mail was stuck.
    In CyberTell: enter what you find under "Dependencies & Outage" as a flow that depends on Exchange Online. Then it shows up in the simulation from step 8.
  3. Save the post-incident review and claim the credit. Service Health keeps reviews for 30 days. SLA credits have to be requested by 31 December.
    In CyberTell: file Microsoft's review and the SLA as documents with the supplier Microsoft under "Suppliers". Your own look back goes into a post-incident review on the incident ticket.
  4. Prepare a channel outside the tenant. A phone tree, SMS, a separate chat or a status page that does not run on Microsoft 365. Send Service Health alerts to at least one address outside it, and keep the important runbooks offline. Test it before you need it.
    In CyberTell: the War room is such a channel, a crisis chat with files and a timeline that runs in Switzerland and does not depend on your Microsoft tenant. Playbooks can be exported as PDF and kept offline.
  5. Never send alarms by email only. Find the building systems, monitoring and sensors that report problems only by email, and give them a second path such as SMS, a phone call or a control room. At Phantasialand, that is exactly where the chlorine warning got stuck.
    In CyberTell: CyberTell does not deliver alarms itself. In the Asset register you record these systems with their criticality and availability, and note in the description which way they raise an alarm.
  6. Do not make recovery depend on mail. No mass password resets during an outage, they end in MFA loops and lockouts once the service is back. Replace one-time codes by email with passkeys or an authenticator app.
    In CyberTell: the Microsoft 365 Assessment reads through Microsoft Graph, with read permissions only, who has not registered MFA yet, and checks Conditional Access and security defaults. An outage playbook records in advance that no passwords are reset during the disruption.
  7. Keep one security path outside the suite. Logs and detection that do not depend on Microsoft 365, for example in a separate SIEM, and a decision made in advance about which controls fail closed.
    In CyberTell: the SOC dashboard shows connections and alarms straight from the firewall, for example Firewalla, pfSense or Sophos, and receives alerts from your own Graylog. Neither depends on Microsoft 365.
  8. Map your dependencies by upstream service. Write down what stops when Exchange Online stops, including Copilot, Defender, Purview and every flow that reads mail, and test the critical functions against a scenario of 24 to 48 hours. This one did not start during Swiss office hours. The next one might.
    In CyberTell: under "Dependencies & Outage" you connect suppliers, assets and business processes and simulate the failure of Microsoft, with the recovery time objective (RTO) and the maximum tolerable period of disruption (MTPD) of each process. A tabletop exercise of the type "System outage / service disruption" walks the team through the 24 to 48 hours.

The notes "In CyberTell" refer to CyberTell, our own platform. Every step works without it.

We will update this post once Microsoft publishes its post-incident review.

How long can your organisation work without Microsoft 365? Pentesting and consulting →