
Photo: Coolcaesar, 2016, via Wikimedia Commons (CC BY-SA 4.0)
For almost exactly a day last week, a large share of the world’s office workers could not open their email. Starting a little after 8 a.m. Pacific time on Thursday, Nov. 12, 2026, Outlook stopped loading mailboxes for most business and school users of Microsoft 365. Teams, Copilot, SharePoint, OneDrive and a list of security and admin tools stumbled along with it. Courts asked people to phone instead of email, hospitals and family doctors switched to backup channels, and the outage-tracking site Downdetector logged more than 130,000 Outlook problem reports at the peak. Microsoft did not declare the problem over until 8:10 a.m. Pacific on Friday, Nov. 13, 23 hours and 50 minutes after it began.
The short answer: It was not a hack, and it was not one mistake. Microsoft’s preliminary post-incident report blames three things that lined up on the same morning: a software defect that had been sitting quietly on its gateway servers since July, a routine shift of Teams traffic onto the same gateway that carries Exchange Online email, and a pool of servers, nearly 14% of capacity, that was already out of service for an unrelated reason. Under the extra load the hidden bug made servers lock up. Traffic that fled the frozen servers landed on healthy ones and froze those too, until more than 80% of the gateway fleet was unresponsive. The fix was simple once found, putting old code back, but every server that came back online was swamped by the backlog of waiting connections, which is why recovery took most of a day.
Microsoft promised a final post-incident report, with a full list of repairs, by today, Nov. 20. That makes this a good moment to go through what the company has disclosed so far, what the technical terms in its report actually mean, why "turn it off and on again" did not work, how this compares with the other giant cloud failures of the past decade, and what businesses can still do, including claiming service credits before the Dec. 31 deadline. Times below are Pacific unless marked otherwise; Microsoft’s own report uses UTC, which was eight hours ahead of Pacific time that week.
What happened, in brief
Microsoft 365 is the subscription bundle that includes Outlook email, Teams chat and meetings, Word, Excel, OneDrive storage, SharePoint sites and, increasingly, the Copilot AI assistant. For most businesses the email part runs on a service called Exchange Online. When your Outlook app, a phone mail app or a browser tab asks for your inbox, the request first passes through a layer of Microsoft servers the company calls the networking gateway. The gateway is the front door: it accepts the connection, negotiates the secure link, and passes the request to whichever back-end system holds your mail.
On Nov. 12 that front door jammed. The mailboxes themselves were fine, and so was the mail inside them. Microsoft’s report says the most significant impact was that "most users" were unable to reach their mailbox through Outlook or other mail clients, and that email delivery was delayed for many. Services that lean on Exchange, including parts of Copilot, Teams, SharePoint, OneDrive, Defender and Purview, were hit as a side effect, though most of them "largely remained online and accessible," according to the report. The full list of affected services in the incident record runs to 14: Exchange, Copilot, Teams, SharePoint, OneDrive, Defender, Purview, Fabric, Intune, Planner, Power Apps, To Do, Power BI and Universal Print.
- When: Thursday, Nov. 12, 8:20 a.m. Pacific (16:20 UTC) to Friday, Nov. 13, 8:10 a.m. Pacific (16:10 UTC), per the preliminary report.
- Incident numbers: EX1550543 for Exchange Online and MO1550821 for the wider Microsoft 365 impact. Administrators see both in the Service health section of the Microsoft 365 admin center.
- Where: Worldwide. The BBC reported problems in at least 74 countries.
- Cause: A latent defect in gateway servers, plus extra load from a Teams failover, plus about 14% of gateway capacity already offline.
- Not the cause: A cyberattack. Microsoft said it was "a technical issue not related to cyber or AI / agents."
- Was mail lost? Microsoft says no. It extended the window for holding undelivered messages to 48 hours, and its report lists as completed the work of making sure queued mail was delivered rather than bounced.
Microsoft’s executive for Copilot, agents and platform, Charles Lamanna, summarized the same three causes on X and apologized. "We are deeply sorry for the impact to all our customers and recognize the impact that this has on their normal business and operations," he wrote, as reported by ITPro.

Photo: Masdestructive, 2009, via Wikimedia Commons (CC BY-SA 3.0)
The hour-by-hour record
Microsoft’s preliminary report includes a detailed log of what its engineers saw and did. Read in order, it is a short story about a team trying one reasonable fix after another while the problem kept moving. The table below converts the report’s UTC times to Pacific and adds three public milestones from Microsoft’s status account and from Downdetector.
| Pacific time | UTC | What happened |
|---|---|---|
| Wed. Nov. 11, 5:07 p.m. | Nov. 12, 01:07 | Automated failover moves Teams traffic onto gateway front-end servers shared with Exchange Online |
| Nov. 11, 5:07 p.m. to Nov. 12, 8:00 a.m. | 01:07 to 16:00 | Load on those servers climbs gradually; isolated server failures in some regions are handled by automation |
| Thu. Nov. 12, 8:03 a.m. | 16:03 | Connectivity anomaly monitor flags a rise in user connection problems |
| 8:17 a.m. | 16:17 | Automation starts redirecting traffic to other healthy endpoints |
| 8:20 a.m. | 16:20 | High-priority alert for Exchange Online; EX1550543 posted to the Service Health Dashboard; official start of impact |
| 8:33 a.m. | 16:33 | First customer support case; reports begin piling up on Downdetector |
| 9:54 a.m. | 17:54 | First post on the Microsoft 365 Status account on X, citing mailbox problems in the Americas |
| 10:07 a.m. | 18:07 | Microsoft says the issue "may affect Microsoft 365 service connectivity for users in any region" |
| About 10 a.m. | About 18:00 | Downdetector Outlook reports peak at 138,058, according to Reuters |
| 10:27 a.m. | 18:27 | MO1550821 posted for Teams, Copilot, Defender and other services |
| 11:00 a.m. | 19:00 | All deployments and configuration changes paused as a precaution |
| 11:20 a.m. | 19:20 | Engineers determine the load is not a traffic surge but servers dropping out and dumping work on the rest |
| 1:00 p.m. | 21:00 | A new health-check method lets one group of servers rejoin; it is rolled out more widely |
| 2:00 p.m. | 22:00 | Telemetry shows the fixes are not working; investigation broadened |
| 3:55 p.m. | 23:55 | Servers returned to service stay healthy at first, then freeze again once traffic arrives |
| 4:30 p.m. | Nov. 13, 00:30 | Connection negotiation identified as the main point of failure |
| 5:53 p.m. | 01:53 | Some automatic traffic redirection is switched off |
| 11:20 p.m. | 07:20 | The fix, reverting the defective code path, is applied to part of the fleet |
| Fri. Nov. 13, midnight | 08:00 | Patched servers confirmed stable; rollout expands to regions entering business hours |
| 3:00 a.m. | 11:00 | Recovery completed across parts of Asia Pacific |
| 3:35 a.m. | 11:35 | Europe returns to healthy levels |
| 6:17 a.m. | 14:17 | Global webmail availability back to expected levels |
| 8:10 a.m. | 16:10 | No active customer impact; monitoring and mail-queue cleanup continue |
| 8:15 p.m. | Nov. 14, 04:15 | Final customer updates issued for both incidents |
Two things stand out in that log. The first is the long fuse: the Teams traffic moved more than 15 hours before anything visibly broke, and Microsoft says isolated server failures were already appearing during that stretch and being patched over by automation. The second is the 15-hour gap between understanding the symptom and applying the cure. By 11:20 a.m. on Thursday the team knew the servers were the problem, not the traffic. The revert that actually fixed things did not start reaching servers until 11:20 p.m.
One small inconsistency is worth flagging. Microsoft’s final status posts on Nov. 14, as captured by several third-party trackers, listed an impact start of 10:05 UTC, or 2:05 a.m. Pacific, more than six hours earlier than the preliminary report’s 16:20 UTC. The report itself does not explain the difference. A separate, earlier Teams problem was reported that morning, and BleepingComputer noted at the time that it was unclear whether the two were related. The preliminary report’s 16:20 UTC start is the one used throughout this article.
Cause one: a bug that waited four months
The heart of the incident is a sentence in Microsoft’s report that is easy to skim past: "a latent defect caused Windows servers to become overwhelmed and unresponsive under certain load conditions." Latent means the defect was present but asleep. It needed a particular set of conditions to wake up, and for months those conditions never arrived.
According to the report, the defect came from a fix Microsoft prepared specifically for its Microsoft 365 networking gateway servers. The company stresses that "neither the underlying defect nor the fix applies to broader Windows deployments," which is its way of saying that your own Windows PC or your company’s Windows servers do not carry the bug. The change began rolling out in July 2026. It went through Microsoft’s safe deployment process, reached every gateway server worldwide over several weeks, and then ran "for more than three months" without causing any trouble.
What it did, when it finally misbehaved, was create lock contention inside an internal Windows process, which exhausted the machine’s CPU and memory. That phrase needs unpacking, because it explains nearly everything that followed.
What lock contention means
Modern servers do many things at once. A gateway server handling tens of thousands of connections runs many threads, small streams of work that each handle a piece of the load, spread across all of the processor’s cores. Sometimes those threads need to touch the same shared piece of data, such as a table of open connections. To stop two threads from changing that data at the same moment and corrupting it, programs use a lock: a thread must grab the lock before touching the data and release it afterward. Everyone else waits.
A lock is like a single-occupancy restroom with a key on a large block of wood. When few people need it, there is no line and nobody notices the key. When many people need it at once, a queue forms, and the time spent waiting can exceed the time spent doing anything useful. In computing terms, that queue is contention. Bad enough contention can be worse than a simple slowdown. Threads waiting for a busy lock may keep spinning on the processor to check whether it is free yet, burning CPU while accomplishing nothing, and the work they were supposed to do piles up in memory. Microsoft says exactly this happened: contention inside the process caused "the machine’s resources (CPU & memory) to be exhausted," and the servers stopped responding.
Microsoft has not said which lock, which process, or what the July change was meant to fix. The preliminary report describes the defect only at that level, and the final report may say more. The report’s later log entries narrow it down somewhat: at 4:30 p.m. Pacific on Thursday, engineers "identified the connection negotiation as the primary area responsible," and determined "that competing processes were preventing connections from being handled successfully." Half an hour later they found "a process on the affected servers responsible for establishing connections was becoming overwhelmed." In other words, the jam sat in the part of the server that sets up new connections, the handshake every Outlook app performs before it can ask for mail.

Photo: Geekerwan, 2023, via Wikimedia Commons (CC BY 3.0)
Why more cores made it worse
The most striking technical detail in the report is about hardware. Microsoft says its load testing missed the problem because of the traffic mix and the server models used in its labs. "The contention issue is worsened on machines with more cores and newer hardware," the report says, "whereas our load testing tended to focus on weaker hardware with fewer cores."
That runs against intuition. A newer, bigger server should be safer, not riskier. But lock contention is one of the classic cases in which more parallel hardware can hurt. If one lock protects a shared resource, only one core can hold it at a time no matter how many cores the machine has. A server with 16 cores might have a handful of threads waiting. A server with 128 cores, which is what a single high-end server chip such as AMD’s EPYC 9754 offers, can have many times more threads competing for the same lock, and the queue grows accordingly. The overhead of handing the lock from core to core also rises on large chips, because the data has to travel farther across the silicon. Engineers sometimes call this a scalability cliff: performance improves with added cores up to a point, then falls off sharply.
The lesson Microsoft draws in the report is narrow but important. A test lab is only useful if it looks like production. If the newest, biggest machines in the fleet are the ones most likely to break, testing on older, smaller machines provides false comfort. Expect the final report’s repair list to include changes to how gateway load tests choose hardware.
Cause two: Teams moved in the night before
Microsoft runs two separate gateway systems for Microsoft 365. The report calls them "two independent and heterogenous network gateway systems designed for isolation and high availability." Heterogeneous means they are built differently, so that a bug in one is unlikely to exist in the other. Services can move between them when one has problems.
At 5:07 p.m. Pacific on Wednesday, Nov. 11, about 15 hours before the outage, Microsoft’s automated failover system moved Teams traffic off the alternative gateway and onto what the report calls the "primary" gateway, the one that also carries Exchange Online. Microsoft describes this kind of move as "part of normal operations and integral to standard Microsoft resiliency efforts." The report does not say why Teams failed over that evening.
The move did two things. It added load to the primary gateway, and it "changed its traffic patterns." The second point matters as much as the first. Teams traffic is not the same as email traffic. Chat clients, meeting clients and presence updates open and hold connections differently from mail clients that sync a mailbox. A different mix of connection types could exercise the vulnerable code in a different way, which fits the report’s mention that the lab’s "traffic mix" also hid the bug. For the next 15 hours the load rose, isolated servers began failing in some regions, and automation quietly cleaned them up.
Cause three: 14% of the servers were already out
The third factor is the plainest. On Nov. 12, "nearly 14% of server capacity for the networking gateway was out of service due to an unrelated issue from the previous day." That capacity "was on track to be returned to active service through routine maintenance and health processes." Microsoft says that on its own, the missing capacity caused no impact, because the service "had sufficient capacity headroom."
Headroom is the spare capacity every large service keeps so that it can lose some machines and still carry peak load. Losing 14% of servers means each remaining server carries about 16% more work than usual, because the same traffic is spread over 86% of the machines. That is normally fine. But stack it on top of extra Teams traffic and the U.S. morning peak, and a server that would have stayed below the bug’s trigger point on an ordinary day could cross it.
This is how many big outages work. No single cause would have done it. The Teams failover was routine. The missing capacity was within the safety margin. The July change had run for months without trouble. Microsoft’s report puts it in one sentence: "Taken together, the increased load and the reduced server capacity, caused the network gateway to encounter the latent bug from July." Safety engineers have a name for this pattern, borrowed from accident investigation: the Swiss cheese model, in which several layers of protection each have holes, and harm happens only when the holes line up.
How one stuck server became 80% of the fleet
Big services are built on the assumption that individual servers will fail all the time. Disks die, power supplies pop, software crashes. So every request is routed by traffic management systems that check whether each server is healthy and steer users away from any that are not. Microsoft’s report calls this "critical to avoiding any interruption due to hardware or software defects and is normally what makes the service resilient."
On Nov. 12 that safety net became the mechanism of collapse. When the first gateway servers locked up, the system did exactly what it was designed to do and moved their users to other servers. But the other servers were running the same code, under the same conditions, with the same latent bug. "The redirected traffic increased the likelihood for the same problem to occur on the remaining endpoints," the report says. "This caused a cascading failure from server to server as the same amount of load was spread across fewer servers."
The arithmetic of a cascade is unforgiving. If a fifth of the servers in a pool fail and their work is spread evenly, each survivor carries a quarter more than before. If half fail, each survivor carries double. If 80% fail, each survivor carries five times its normal load. And every server that tips over makes the next one more likely to tip. Microsoft says that "in less than two hours, the previously unknown defect was encountered across our fleet," and that "at its lowest, more than 80% of servers were unresponsive."
"At its lowest, more than 80% of servers were unresponsive." Microsoft’s preliminary post-incident report, Nov. 15, 2026
There is a second multiplier the report alludes to in its timeline: clients retry. When Outlook cannot connect, it tries again. So do phone mail apps, sync engines, scripts and connected business software. Every failed attempt becomes several more attempts. Prof. James Ferryman of the University of Reading described the dynamic to the BBC while the outage was still under way: "One server falls over, its work gets passed to the next one, and meanwhile huge numbers of phones and laptops keep trying to get back in."

Photo: Chris Heaton, 2023, via Wikimedia Commons (CC BY-SA 2.0)
Engineers have a vivid name for the moment a fleet of clients all pile back in at once: a thundering herd. Google’s own Site Reliability Engineering book devotes a chapter to cascading failures and warns that retries, if not limited and spread out, can turn a small overload into a self-sustaining one. Amazon’s guidance for its engineers recommends what it calls exponential backoff with jitter, which means each client waits longer after every failure and adds a random delay so millions of devices do not all retry in the same second. Whether Microsoft’s clients and gateway handled backoff well is one of the questions the final report may address.
Why the bulkheads leaked
Large cloud services are not supposed to fail globally. The standard defense is partitioning: split the service into many independent cells, so that a problem in one stays in that one. Microsoft says its gateway is built that way. It is "composed of thousands of ‘capacity units’ which represent shards of capacity, hardware, and networking infrastructure," and those units "are then organized into regions."
The term engineers use for the walls between those cells is bulkheads, borrowed from shipbuilding. A ship’s hull is divided by watertight walls so that a hole in one compartment floods that compartment and not the whole ship. In software, a bulkhead stops one cell’s overload from spilling into its neighbors, for example by refusing to accept another region’s redirected traffic beyond a fixed limit.

Illustration: Seaman’s Pocket-Book, 1943, via Wikimedia Commons (Public domain)
On Nov. 12 the bulkheads did not hold. "In this case, due to a separate issue, the normal bulkheads between these different partitions did not stop the cascading failure," the report says. "This is what allowed the impact to spread across our network gateway infrastructure globally." Microsoft has not yet explained what that separate issue was. It is arguably the most important unanswered question in the preliminary report, because it is the difference between a regional outage that inconveniences one continent for a few hours and a global one that halts email in dozens of countries for a day. One of the completed repair items in the preliminary report, "implemented guardrails for traffic distribution to avoid cascading impact," appears aimed squarely at it.
The software engineer and newsletter writer Gergely Orosz, a former engineer at Skype and Uber, made the same point to the BBC, suggesting that a solution could be to split services per region to limit the impact, as Amazon does with its cloud regions. Microsoft’s design already includes regions. The trouble was that, this time, the walls between them let the water through.
Why Teams mostly worked and email did not
If you were in the middle of the outage, you may have noticed something odd: you could not get your email, but Teams chats and meetings mostly worked. The report explains why. Services that were able to "failed out of the faulty network gateway infrastructure," naming Teams, OneDrive, SharePoint and Office apps. That is why those services "were partially restored earlier, with Teams chats and meetings functioning without issue for most of the outage."
Exchange Online could not do the same. "Not all services, like Exchange Online, can fail out of this network gateway for scale reasons," the report says. In plain terms, the alternative gateway is not built to carry the full weight of the world’s business email on top of everything else. Email clients are numerous and chatty: phones, laptops and mail apps keep connections open or reconnect frequently to check for new messages, all day, for every business user. Moving that load onto a second system that was never sized for it would risk taking that system down too.
This is an uncomfortable detail for anyone who assumed Microsoft’s two-gateway design protected email. It protected email from failures in the other gateway. It did not give email a second road to travel when its own road was blocked. Any serious repair plan probably has to change that, either by building the alternative gateway big enough to take Exchange in an emergency or by splitting Exchange traffic so that a portion of it can move.

Photo: PiDatacenters, 2021, via Wikimedia Commons (CC BY-SA 4.0)
Why fixing it took nearly a day
The most common question during the outage was why a company with Microsoft’s resources could not just roll back whatever broke. The report gives three reasons, and together they explain the slow recovery better than any outside commentary did.
First, there was nothing recent to roll back. Most outages at big cloud providers follow a fresh change, and the standard response is to undo it. Microsoft’s tools for this, which it lists as "rollback, flighting, and feature flags," are designed for recent changes. Flighting is Microsoft’s term for exposing a change to a small group first; a feature flag is a switch that turns a code path on or off without redeploying software. But this defect had been in place since July. "Since the main code defect was a latent bug from several months previous, our normal rollback, flighting, and feature flags were unable to quickly revert the change," the report says. Microsoft’s early status posts reflected this: the company said at one point it had found no recent change tied to the start of the problems, according to the San Francisco Standard’s coverage.
Second, engineers spent hours treating symptoms. The timeline shows a team working through a reasonable sequence. They suspected bad health checks were knocking good servers out of service, so at noon Pacific they deployed a change to bypass the health-check components that were failing. At 1 p.m. they tried an alternative health check method, which let some servers rejoin. They built automation to keep restarting servers that stopped responding. They throttled traffic, adjusted protocols and blocked some traffic outright. Each step helped somewhere and held nowhere. The 4:30 p.m. Pacific finding that connection negotiation was the choke point led them toward the real defect, and the actual fix, an update that "returned the connection handling code path to the previously known good state," began reaching servers at 11:20 p.m.
Third, the fix itself could not be applied the usual way. "Recovery took longer than usual because as servers returned to service, they were immediately overwhelmed by waiting traffic and became unresponsive again," the report says. Microsoft’s "normally automated recovery processes were not designed to restore service from this broad unresponsive state." Its 3:55 p.m. log entry captures the trap: servers returned to service "initially remained healthy but then become unresponsive again upon receiving traffic." So the team took groups of servers out of rotation, applied the change, restarted them and put them back, working through "large quantities of infrastructure at a time (e.g. 50%+ of a region)." They prioritized regions that were entering their business day, which is why Asia Pacific and then Europe recovered first in the early hours of Friday, Pacific time, and why North America’s webmail was back by 6:17 a.m.
A Microsoft spokesperson defended the pace to the BBC: "given the scale of demand" it restored services gradually, and "while this took longer, it helped ensure a stable restoration." Orosz, speaking to the BBC, said he could not recall an outage of business email ever taking this long to return to normal.
The trap has a name: metastable failure
Computer scientists have studied this exact pattern. In a 2021 paper titled "Metastable Failures in Distributed Systems," researchers including Nathan Bronson described systems that can sit in two stable states: a healthy one, and a broken one that keeps itself going even after the original trigger is gone. A trigger such as a load spike pushes the system into the bad state, and a sustaining effect, often retries or the cost of recovering, keeps it there. A follow-up study presented at the OSDI conference in 2022, "Metastable Failures in the Wild," combed public incident reports and examined 22 such failures at 11 organizations. It found the pattern "universally observed" and estimated that at least 4 of 15 major Amazon Web Services outages over the preceding decade were metastable failures.
The Microsoft outage reads like a textbook case. The trigger was the bug meeting extra load. The sustaining effect was the backlog: every recovered server faced not normal traffic but the pent-up demand of millions of devices that had been retrying for hours. Turning off the original trigger was not enough. Microsoft had to keep the load away while it fixed the servers, then let traffic back in gradually, which is exactly what the report describes. Amazon’s own postmortem of its October 2025 outage described a similar state in one of its internal systems, which it called "congestive collapse," and Google said its large us-central1 region took up to about 2 hours and 40 minutes to recover in June 2025 because restarting tasks created "a herd effect" on a database they all depended on.
Who felt it, and how much
Microsoft has not said how many users were affected, and the BBC noted that the full scale "has not been confirmed." What is known comes from outside measurements and from organizations that went public with their problems.
Downdetector, which counts user-submitted problem reports, saw Outlook reports climb from a few thousand to a peak of 138,058 around 1 p.m. Eastern, or 10 a.m. Pacific, according to Reuters; other outlets that sampled the site at slightly different moments recorded about 131,000. By evening the Outlook count had fallen to around 34,490, Reuters reported. Those numbers count complaints, not people. One frustrated worker may report twice, while a company of 5,000 employees may generate no reports at all, so the true number of affected users was certainly far higher. The BBC reported that more than 430 million people use Microsoft 365 apps, citing Microsoft’s website, and that problems were reported in at least 74 countries.
The disruption reached well beyond private offices. The BBC reported that courts from Los Angeles to Texas warned that communications might be delayed, that a healthcare network running several Toronto hospitals had email trouble while patient care was unaffected, that family doctors’ surgeries in the U.K. asked patients not to send messages through their websites, and that Japan’s government said its shared IT platform was disrupted. In San Francisco, city departments struggled to communicate, according to the San Francisco Standard, and the city’s Superior Court told the public on Friday that "court staff may be delayed in receiving or responding to emails." OpenAI said an issue with its connections to Microsoft services had caused some of its tools not to work as expected.
Not everyone was down. Users of personal Outlook.com accounts were not clearly covered by the incident: Microsoft’s report concerns its work and school services and does not say whether consumer accounts were affected. Rival platforms were untouched. Google Workspace, Amazon Web Services and Cloudflare reported no related problems, according to trackers that followed the event.
What happened to the email
For many people the bigger worry was not the lost day but lost messages. Here Microsoft’s design held up. Internet email is built to be patient: when a receiving server is unreachable or answers with a temporary error, the sending server keeps the message and tries again later, typically for days. Microsoft said early Friday that it had extended its "email timeout window" to 48 hours "to help protect delayed messages as delivery catches up," and the preliminary report lists as completed the work of making sure "delayed messages weren’t returned as undeliverable and all messages that were queued for delivery during the event were processed."
Two side effects are worth knowing. Delayed messages arrived with the time they were delivered, not the time they were sent, as the BBC noted, so an email that appears to have come in Friday morning may have been sent Thursday afternoon. And while email was delayed, the security checks that ride along with it were also stressed: trackers reported that Safe Links, the Defender feature that checks URLs in messages when you click them, may have been unable to verify some links during the event.

Photo: Robert Scoble, 2010, via Wikimedia Commons (CC BY 2.0)
The phishing that followed
Criminals moved quickly. Microsoft Threat Intelligence said it saw a phishing campaign that exploited the disruption, with fake messages posing as Microsoft 365 service health notices and even citing the real incident number, EX1550543, according to SecurityWeek. Some messages detected on Nov. 13 falsely told recipients that email held during the outage would be deleted within 24 hours unless they clicked a link to restore it. Subject lines began with "Outage -" followed by a six-character code. Microsoft described the activity as "consistent with credential phishing" and later said its security products were blocking it.
The general rule applies: Microsoft does not ask users to click a link in an email to rescue messages after an outage. Real incident information appears in the Service health section of the Microsoft 365 admin center, which only administrators can see, and on Microsoft’s public status page and status account on X.
What Microsoft has fixed so far
The preliminary report closes with a short table of repair items. Five are marked as completed:
- An update that "returned the connection handling code path to the previously known good state," removing the July defect.
- An extended email delivery window so queued mail was delivered rather than bounced.
- Additional server capacity, added "as a proactive measure to help with recovery."
- Additional monitoring for traffic-redirection patterns and connection-handling health.
- "Guardrails for traffic distribution to avoid cascading impact."
The sixth line reads: "A comprehensive list of repair items will be delivered as part of the final PIR," with a date of 11/20/2026. Microsoft also says that "many longer-term changes from our full root cause analysis will be in the final Post Incident Report." The preliminary report itself says a final report will be provided within five business days of full resolution and will supersede the preliminary one.
Microsoft distributes these reports to administrators through the admin center rather than on a public web page, which is why most public coverage, including this article, relies on copies that administrators and news sites reproduced. As of Friday morning, Nov. 20, a final report had not appeared publicly. Administrators should check MO1550821 and EX1550543 in Service health and save a copy, because the admin center keeps service health history for only a limited time.
What the final report needs to answer
Based on the preliminary report’s own gaps, these are the questions to look for:
- What failed in the bulkheads? The "separate issue" that let a regional problem go global is the single most important unexplained detail.
- Why could Exchange not fail over, and will Microsoft build a path that lets at least part of its email traffic move to the alternative gateway?
- What took 15 hours? Engineers knew by 11:20 a.m. Pacific that unhealthy servers, not traffic, were driving the load. The revert did not start until 11:20 p.m.
- What was the July change, and how will load testing change to cover newer, many-core servers and realistic traffic mixes?
- What were the earlier failures? The report says isolated server failures began in some regions during the 15 hours after the Teams failover and were handled by automation. Were they early warnings that went unread?
- What took 14% of capacity out the day before?
- Why did the start times differ between the report and the final status posts?

Photo: Tedder, 2025, via Wikimedia Commons (CC BY 4.0)
How it compares with other giant outages
Every few years a single failure at a large technology company knocks out services for a large share of the internet’s users. The Microsoft 365 outage belongs in that list, and the comparison is useful because the causes rhyme. The table below draws on each company’s own published postmortem where one exists.
| Outage | Date | What triggered it | How long |
|---|---|---|---|
| Microsoft 365 gateway | Nov. 12-13, 2026 | Latent July defect, Teams failover load and 14% of capacity already offline | 23 hours 50 minutes |
| Cloudflare | Nov. 18, 2025 | Database permissions change doubled the size of a Bot Management feature file | About 3 hours of core impact; all systems normal after nearly 6 hours |
| Amazon Web Services, us-east-1 | Oct. 19-20, 2025 | Latent race condition left DynamoDB’s regional DNS record empty | About 14.5 hours, in three phases |
| Google Cloud | June 12, 2025 | New quota-check code shipped May 29 without a feature flag crashed on blank policy fields | About 7.5 hours for the full incident |
| CrowdStrike | July 19, 2024 | Faulty security sensor content update crashed Windows PCs | 8.5 million devices; recovery took days for many organizations |
| Oct. 4, 2021 | Maintenance command took down the backbone; an audit-tool bug failed to stop it | Roughly six hours | |
| Fastly | June 8, 2021 | Valid customer configuration triggered a bug deployed May 12 | 95% of network normal within 49 minutes |
| Amazon S3, us-east-1 | Feb. 28, 2017 | Mistyped command removed more servers than intended | About 4 hours 15 minutes |
A few patterns jump out. Latent bugs are a recurring theme. Fastly’s bug was deployed on May 12, 2021 and sat dormant until a customer’s perfectly valid configuration triggered it on June 8, when "85% of our network" returned errors, the company wrote. Google’s June 2025 failure came from code shipped on May 29 that waited for a particular kind of policy data to arrive. Amazon’s October 2025 outage stemmed from "a latent race condition in the DynamoDB DNS management system." Microsoft’s July defect waited longer than any of them.
Recovery, not the initial failure, is often what makes an outage long. Cloudflare had core traffic "largely flowing as normal by 14:30" UTC, a little over three hours after it started, but spent the next few hours dealing with "increased load on various parts of our network as traffic rushed back online," and declared all systems normal at 17:06 UTC. Google found that restarting tasks in its biggest region overloaded the database they all read from, because the software "did not have the appropriate randomized exponential backoff implemented." Amazon described an internal system that "had entered a state of congestive collapse." Microsoft’s report describes the same family of problem at a larger scale: servers that died again the moment traffic found them.
Then there is the matter of reach. Microsoft’s July 2024 estimate that the CrowdStrike update affected 8.5 million Windows devices, "less than one percent of all Windows machines," shows how a small percentage can still stop airlines and hospitals when it lands on the right machines. The Microsoft 365 outage was different in kind: it did not break customers’ computers, but it took away a service that a vast number of organizations treat as their main communication channel. For a day, many of them had no fallback ready.
Readers who followed aglena’s coverage this week will see a theme. The Apple outage on Nov. 17 and the Chase outage on Nov. 16 were shorter, but each showed how much daily life now depends on a handful of central systems.

Photo: Johan Wessman / News Øresund, 2021, via Wikimedia Commons (CC BY 2.0)
Service credits: what businesses can claim, and by when
Microsoft’s Service Level Agreement for Online Services promises business customers a level of uptime and pays service credits, not cash, when it misses. For Exchange Online, downtime is defined as "any period of time when users are unable to send or receive email with Outlook Web Access," the browser version of Outlook. Uptime is calculated per tenant in user-minutes: the total minutes in the month multiplied by the number of users, minus the minutes each affected user lost.
| Monthly Exchange Online uptime | Service credit |
|---|---|
| Below 99.9% | 25% of that month’s Exchange Online fees |
| Below 99% | 50% |
| Below 95% | 100% |
November has 43,200 minutes. The outage lasted 1,430 minutes by Microsoft’s own start and end times. If every user in a company lost Outlook on the web for the whole window, that company’s November uptime would be about 96.7%, which falls in the 50% credit tier. If only half the users were affected for the whole period, uptime would be about 98.3%, still in the 50% tier. A company whose users lost access for only six hours would land at about 99.2%, in the 25% tier. These are illustrations; the real figure depends on each tenant’s measured downtime, and Microsoft decides whether a credit is owed after a claim is filed.
Three caveats matter. First, credits are not automatic. Administrators must file a claim with Microsoft support with the details the SLA requires, including a description of the incident, the time and duration of downtime, affected resources, the number and location of affected users, and the errors they saw. Second, credits apply only to the service that missed its target, so an Exchange Online outage earns credits against Exchange Online fees, not the whole Microsoft 365 bill. Third, there is a deadline. For services other than Azure, the SLA says Microsoft "must receive the claim by the end of the Applicable Period following the month in which the Incident occurred." For a November incident, that means Dec. 31, 2026. Customers who buy through a reseller or Microsoft partner usually file through that partner.
The SLA also says credits are the "sole and exclusive remedy" for availability problems and do not compensate for lost revenue or other indirect losses. For a business that lost a day of sales, half a month of Exchange Online fees is not much. It is still worth claiming, and the paper trail is worth keeping for insurers and for any future contract talks.
What IT teams should do now
For administrators, the outage is over, but a few tasks are worth finishing while the details are fresh:
- Save the reports. Download the preliminary and final post-incident reports from Service health under MO1550821 and EX1550543.
- Check mail flow. Use message trace in the Exchange admin center to confirm that messages sent and received during Nov. 12-13 were delivered. Microsoft says queued mail was processed, but business-critical messages deserve a check.
- File the credit claim before Dec. 31, with downtime per user and screenshots or logs of errors.
- Warn staff about the phishing that impersonated outage notices, and remind them that real outage information comes from IT, not from a link in an email.
- Write down what broke. Which processes stalled because email was down? Approvals, password resets, two-factor codes sent by email, vendor orders, ticket systems? That list is the start of a continuity plan.
- Set up a backup channel that does not depend on the same provider: a phone tree, a text-message alert service, or a secondary chat or email system for a small group of key staff.
- Know where to look. If the admin center itself is unreachable, Microsoft’s public status page and the Microsoft 365 Status account on X carry updates. Bookmark both.

Photo: NASA/Robert Markowitz, 2026, public domain
What ordinary users can do next time
For individual users there is less to do, but a few habits help:
- Do not keep resending. During an outage, outgoing mail usually sits in your outbox or in a queue and will go out when service returns. Resending creates duplicates.
- Check the official status before assuming the problem is yours. Rebooting your laptop or reinstalling Outlook will not fix a server-side outage.
- Keep important contacts and documents available offline. The desktop Outlook app keeps a local copy of recent mail, and OneDrive can keep chosen files on your device. Both help when the service cannot be reached.
- Be suspicious of outage emails. Any message urging you to click a link to "restore" mail after an outage should be treated as phishing.
- Expect odd timestamps. Messages delayed by an outage show the time they arrived, not the time they were sent.
What it says about how the cloud is built
Microsoft has written publicly for years about safe deployment, its practice of shipping changes in stages, first to test environments, then to small slices of production, then wider, with health monitoring at every step so a bad change can be caught and stopped early. In a February 2020 post on the Azure blog, Azure’s chief technology officer, Mark Russinovich, described the approach in detail: changes go through test and integration checks, then into "canary regions" that are full Azure regions used for early updates, and then out to production "in a gradual manner," with health signals measured continuously "so that we can detect in relative isolation if there is any unexpected impact associated with the change that did not surface during testing."
The Nov. 12 outage did not show a failure of that process in the usual sense. The July change went through it and passed, because the bug did not show up under the conditions it met during rollout. That is the uncomfortable lesson: staged deployment catches bugs that appear soon. It cannot catch a defect that only appears when a rare combination of load, traffic mix and hardware lines up months later. For those, the defenses are different. They include testing on hardware that matches the newest production servers, load tests that push far beyond normal peaks with realistic traffic, bulkheads that hold even when automation is trying to move traffic across them, and recovery procedures that are practiced for the worst case, when most of a fleet is down at once.

Photo: Derrick Coetzee, 2011, via Wikimedia Commons (CC0)
There is a second lesson, about the automation itself. Microsoft’s traffic management did its job too well. It saw servers failing and moved their load elsewhere, which spread the failure. At 5:53 p.m. Pacific on Thursday the team "disabled some automated traffic redirection mechanisms," turning off the system that was supposed to help. Large providers have learned this before. Amazon’s builders’ library describes a design idea it calls static stability, in which "the overall system keeps working even when a dependency becomes impaired," without needing to make big changes in the middle of a failure. Automation that reacts to failure by shifting load must also know when to stop, or it becomes an amplifier.
The third lesson is about concentration, and it is the one the BBC’s interviewees kept returning to. When one company’s gateway jams, courts, clinics, schools and companies in dozens of countries lose email at the same time. Walter Pasquarelli, an AI strategist, told the BBC that "as AI assistants are plugged into our email, calendars and files, a failure like this may spread even further." Microsoft rejected any link between this outage and AI. Its status update called it "a technical issue not related to cyber or AI / agents." But the point about dependence stands regardless of the cause.
What we know, and what is still unclear
Established by Microsoft’s preliminary report:
- The outage ran from 8:20 a.m. Pacific on Nov. 12 to 8:10 a.m. Pacific on Nov. 13 (16:20 to 16:10 UTC).
- The trigger was a latent lock-contention defect in gateway servers, introduced by a change that began deploying in July.
- A Teams failover onto the shared gateway, about 15 hours earlier, raised load and changed traffic patterns.
- Nearly 14% of gateway capacity was offline because of an unrelated issue from the previous day.
- More than 80% of gateway servers were unresponsive at the worst point.
- The fix reverted the connection-handling code to its earlier state, applied group by group.
- Exchange Online could not fail over to the alternative gateway "for scale reasons."
Still unclear:
- Exactly how many users and organizations were affected.
- What "separate issue" prevented the bulkheads from containing the failure.
- Why Teams failed over the evening before, and what took 14% of capacity out of service.
- What longer-term changes Microsoft will commit to in the final report.
- Why Microsoft’s final status posts gave an earlier start time than the preliminary report.
Questions and answers
Was it a cyberattack?
No. Microsoft said it was "a technical issue not related to cyber or AI / agents," and its status updates during the outage said there was "no indication this is the result of malicious activity." The phishing that followed exploited the outage; it did not cause it.
Did AI or Copilot cause the outage?
Microsoft says no, and nothing in the preliminary report points to AI. The defect was in connection-handling code on gateway servers. Copilot was among the services affected because parts of it rely on Exchange Online.
Was any email lost?
Microsoft says delayed mail was kept and delivered, not bounced, and that it extended the delivery window to 48 hours. If an important message never arrived, ask your administrator to run a message trace.
Why did Teams work when Outlook did not?
Teams, OneDrive, SharePoint and Office apps were able to move off the broken gateway. Exchange Online, which serves Outlook, could not, because of its scale.
Was my home Windows PC affected?
No. Microsoft says the defective fix was made specifically for its Microsoft 365 gateway servers and does not apply to broader Windows deployments. The outage was on Microsoft’s side.
Can my company get money back?
Business customers may qualify for service credits under Microsoft’s SLA, likely in the 50% tier for Exchange Online if users were affected for most of the outage. Claims must reach Microsoft by Dec. 31, 2026.
Was this Microsoft’s worst outage?
Ars Technica called it one of the largest outages in the company’s history, and Orosz told the BBC it could be the "most disruptive" business email outage "of all time." Microsoft has not ranked it.
Could it happen again?
Some version of it could happen to any large provider. The specific defect has been removed, and Microsoft has added guardrails on traffic distribution and new monitoring. Whether the deeper weaknesses, the leaky bulkheads and email’s lack of an alternate route, are fixed will depend on what the final report commits to and how quickly it is delivered.
What happens next
Microsoft’s final post-incident report is due today, Nov. 20, through the admin center. It should include the "comprehensive list of repair items" the preliminary version promised. Watch for specific commitments on bulkhead isolation, an emergency failover path for Exchange, load testing on newer hardware, and limits on how far automation can push traffic during a failure. Administrators have until Dec. 31 to file service credit claims for the November incident. And large customers, especially in government and healthcare, are likely to ask Microsoft harder questions about continuity plans in their next contract reviews.
The bottom line
The Microsoft 365 outage of Nov. 12-13, 2026 was a failure of combination, not of one bad decision. A bug that had passed every test and run quietly for four months met an ordinary traffic shift and an ordinary capacity shortfall during the busiest hours of the U.S. workday. Automation designed to protect users spread the failure, partitions designed to contain it did not, and the very backlog of people trying to get their email back kept knocking the servers down again. Microsoft’s preliminary report is unusually candid about all of this. The final report will show whether the company has found ways to make sure that the next latent bug, which almost certainly already exists somewhere, cannot take the world’s email down with it.
Further reading

Site Reliability Engineering: How Google Runs Production Systems
Google’s own engineers explain how they keep huge services running, including a full chapter on cascading failures and the retry storms that stretched Microsoft’s recovery.

Release It! Design and Deploy Production-Ready Software
Michael Nygard’s practical guide to the stability patterns at the center of this outage, from bulkheads and circuit breakers to timeouts and load shedding.

APC Back-UPS BE600M1 Battery Backup and Surge Protector
A small, popular battery backup that keeps a modem, router and desktop running through short power cuts, the local half of staying online when something fails.

Logitech C920x HD Pro Webcam
Teams meetings kept working for most of the outage; this long-running 1080p webcam is still one of the most widely used for video calls.

Amazon Echo Dot
An inexpensive smart speaker that runs on a different cloud, handy for timers, reminders and news headlines when work tools are down.
