Every engineer has lived the moment where the escalation path runs through the thing that's broken. The pager is on the same cloud as the incident. The runbook is in the wiki that's down. This week, a very large version of that played out in inboxes everywhere.
The outage was still partly ongoing as we wrote this, so details may change.
What happened
A Microsoft 365 and Exchange Online outage began around 16:00 GMT on 12 November 2026. Within about an hour, users in 74 countries were reporting problems, and it ran past 12 hours across Outlook, Teams and Copilot. CNN kept a live blog as it unfolded, and Gizmodo noted services were still unstable late into the evening.
Outlook reports on Downdetector peaked at roughly 130,000 to 140,000, about 2,000 times normal.
The status updates are where it got interesting:
- Early updates blamed "elevated load" across multiple regions and said traffic was being redistributed.
- The first posts said the problem was "in the Americas" and were later corrected to say any region.
- Microsoft's stated mechanism was a process on the affected servers, responsible for establishing secure connections, that becomes overwhelmed and stops processing requests.
- Recovery was flaky. Servers restarted, took traffic, then became unresponsive again. The wording moved from "positive results" to a change "not taking as expected" while the team looked for "alternate workstreams", and then to "marginal success". CBC and Yahoo News UK both led with the fixes not working as expected.
- Microsoft extended the email timeout window to 48 hours to protect delayed messages.
Microsoft says it isn't a cyberattack. It was rolling back a network change, but hasn't said whether that change was to blame, and no root cause has been published. Al Jazeera's explainer and ABC News cover the "progress toward recovery" phase. Wired and Australia's Home Affairs Minister compared it to CrowdStrike, as a bad update gone awry rather than an attack. The OpenAI and Anthropic Microsoft connectors were also disrupted.
Why engineers get it
Nobody is laughing at the people fixing this. Rolling back a network change on a global service while traffic keeps arriving is hard, and nobody enjoys it.
The pain we recognise is the language. "Not taking as expected" is the most polite sentence in operations. It means: we did a thing, we watched a graph, and the graph did not do the thing. "Alternate workstreams" means plan B, and "marginal success" is a green tick with an asterisk. Anyone who has written a status update at 3 a.m. knows how many drafts it takes to say "it's still broken" in a calm voice.
Then there's the flapping. A server comes back, takes traffic, and falls over again. Any engineer who has watched a service heal and immediately get trampled by its own retry queue knows that feeling.
The everyday version is just as familiar. Someone's manager says "just call them", and the support route turns out to be an email address on the platform that's down. The whole office is refreshing an inbox that is, technically, working as designed: waiting.
The design
The design is a two-panel comic. A sweaty engineer holds a landline to their ear while a manager points over their shoulder, saying "Just call Microsoft." In panel two, a phone tree announces that your callback ticket will be sent by email. Behind the engineer, a laptop shows an empty inbox with a spinning wheel.
To be clear, it's a joke about a loop, not a claim about how Microsoft support works. The only path to help runs through the exact system that's down, and that's a bug every one of us has shipped a version of.
It's on mugs and tees for anyone who spent the day refreshing.
The Cover-Up Tour: Nothing to See Here
Git Push and Pray