
Microsoft, FAA and Chase Outages: What We Know
What happened during the November 2026 outages, which failures were connected, and what we know about AI's possible involvement.
Email, flights, banking, hotel accounts: the outages of the past week reached enough of daily life to make the timing unsettling. Between November 12 and 17, failures repeatedly interrupted services people count on. We found disruptions involving ten organizations, although that count includes multiple incidents at Claude and services affected by another provider's failure. The AWS incident was regional, while Microsoft's reached users worldwide.
The evidence supports concern about how essential services handle failures. It does not establish a coordinated AI attack behind this cluster of outages, but AI may be adding pressure in a less obvious way.
November's Major Tech Outages at a Glance
This table reflects public information available November 17. A provider's explanation is attributed where it supplies the account; an undisclosed cause remains unknown.
Organization and Dates | What Stopped Working | Reported Explanation |
|---|---|---|
Microsoft 365, Nov. 12-14 | Outlook and features in Teams, Copilot and other services. | A July software defect surfaced under changed traffic and reduced capacity; automatic recovery spread the failure. Active impact ended Nov. 13, with monitoring continuing into Nov. 14. Preliminary report. |
FAA, Nov. 13 | Network instability prompted a nationwide stop to departures, disrupting travel into the weekend. | Excessive connection attempts overwhelmed older equipment. Isolating two sites stabilized the network; root-cause investigation continues. FAA account. |
Workday, Nov. 13 | Affected customers lost access to workplace systems during emergency maintenance. | Backup-environment services unexpectedly connected to live databases during maintenance validation, leaving data inconsistent. Vendor alert reproduced by NSHE. |
Anthropic/Claude, Nov. 12-13 | Microsoft-connected mail/calendar tools failed; separate login and broader Claude, API and Code incidents occurred Nov. 13. | Microsoft caused the connector disruption. Anthropic's incident entries didn't disclose root causes for the separate Claude failures. Incident history. |
AWS, Nov. 15 | Regional connectivity and sign-in problems in Northern Virginia, resolved that morning. | The reviewed public notice described the affected services without establishing a root cause. AWS event. |
Chase, Nov. 16 | Transactions slowed, including payments involving customers at other banks. | A software bug affected internal hardware systems. Chase said it resolved the problem later Monday and excluded a cyberattack. Chase's explanation to AP. |
United Airlines, Nov. 16 | Departing flights were delayed during an outage lasting more than 90 minutes. | United identified a technology problem and said it found no indication of a cybersecurity incident; a detailed cause wasn't published in the reviewed account. AP reporting. |
Apple, Nov. 17 | Account access, store systems and services were disrupted. | Apple blamed an internal update and said it wasn't a cyberattack. Apple's statement, reported by MacRumors. |
Marriott, Nov. 17 | Marriott.com and Bonvoy app sign-ins failed. | Marriott said it resolved a technical issue and found no indication of a cybersecurity cause. Marriott's statement to AP. |
NSA, Nov. 17 | Some internal systems experienced disruption; the agency said mission-critical capabilities remained operational. | Temporary degradation during scheduled maintenance; a detailed technical cause wasn't supplied. NSA statement. |
Why These Outages Spread So Widely
Microsoft offers the clearest example of a failure becoming much bigger than the defect that started it. Its software problem had been present for months before a change in traffic exposed it. Servers stopped responding, so recovery systems sent their traffic to working machines. Those machines encountered the same defect under the added load and failed too. The system kept shifting work onto a shrinking pool of servers until the disruption spread globally.
That is a serious reliability problem even without an attacker. It also shows why having equipment in different locations doesn't necessarily protect a service: the machines can share the same faulty software. At the FAA, investigators similarly described a network where excessive traffic destabilized older equipment across multiple facilities. Some radio frequencies and radar feeds remained available, and aircraft already airborne landed safely, according to the agency. The ground stop protected safety while technicians stabilized the network. Microsoft report, FAA account.
Chase illustrates another way trouble travels. A payment involves more than one institution, so a customer at a functioning bank can still experience a delay when the other bank has a problem. Zelle's operator said its network remained operational. Treating every bank mentioned on an outage tracker as a separate failure would exaggerate what happened and obscure the actual connection. AP's banking explainer.
Former U.S. deputy chief technology officer Jennifer Pahlka adds an organizational explanation in Derek Thompson's Plain English episode, 'Why Everything Is Breaking All at Once'. She argues that agencies need enough technical expertise in-house to direct vendors, test changes throughout a project and learn from failures. Replacing old equipment alone won't deliver that capability; the people responsible for a service also need the knowledge and authority to improve it.
How AI Could Have Contributed
Microsoft explicitly excluded AI agents as the cause of its outage, and the FAA excluded its SMART AI system. Those statements address particular explanations for the failures.
The established AI connection here is partly dependence: Claude's Microsoft tools failed because Microsoft failed. An AI assistant still needs working accounts, networks and outside services. Adding AI doesn't remove those dependencies, and an outage affecting an AI product tells us little about what caused it. Anthropic's incident record.
There is also a possible indirect connection that officials have raised publicly. Senator Mark Warner said AI is finding vulnerabilities across many systems quickly, creating pressure to install security patches immediately and straining older infrastructure. His assessment points toward security work going wrong as organizations respond to AI-discovered weaknesses. That gives the idea more weight than an unsupported internet theory, although it doesn't establish the chain of events for every outage. Warner's statement, reported by AP.
Cybersecurity journalist Nicole Perlroth gives that pressure a clearer shape in the same episode. Her argument is that AI has accelerated finding and exploiting weaknesses much more than it has simplified repairing them. A flaw may be found quickly, but a safe fix still has to work across different systems without disrupting the services attached to them. That mismatch can leave organizations choosing between the risks of waiting and the risks of rushing an update.
Perlroth also points to the approaching holidays as another source of pressure, with organizations trying to get updates installed before the break. That supports our suspicion that the calendar may be compressing an already difficult job, though it doesn't establish the update schedule behind each outage.
Other possibilities remain speculation: a developer might accept flawed AI-generated code, or an authorized agent could make an unsafe configuration change. Establishing either would require evidence connecting that action to the failure. Conventional automation can also move traffic, apply settings and repeat mistakes, so finding an automated process in the chain wouldn't settle the AI question.
Our concern is that essential services can fail widely during changes, maintenance and recovery, with consequences extending well beyond the organization where trouble begins. The outages justify demanding better explanations and evidence that the failures have been contained. Our banks, airlines and technology providers need to show they can make essential repairs without repeatedly leaving people unable to work, move their money or get home even in a world where AI may make cyber attacks easier and require faster update cycles.

Additional Reading on AI in Cyber Security
A convincing storefront, attractive product photos and a page full of happy reviews can now be assembled with AI. Malwarebytes describes how scammers combine those elements with personalized messages to make fraudulent shops look established, just as holiday buying picks up. Its report also identifies Amazon and major delivery companies as frequent impersonation targets. Before entering payment details, check the seller independently and open the retailer's official site directly instead of following an unexpected offer.
Scammers put an AI-generated Elon Musk into a fake Tesla livestream and promised to double the cryptocurrency viewers sent them. Coinbase says its AI defenses stopped $4.7 million in attempted transfers, about 91% of the money its customers tried to send during the scam. That is a useful example of AI working on the defensive side, too. Any giveaway requiring you to send money before receiving more should end the conversation, however convincing the person on screen looks.
Treasury's annual holiday scam advisory puts impersonation near the center of its warning: criminals can use synthetic voices and video to pose as trusted businesses or loved ones. The agency recommends ending suspicious conversations and contacting the person or institution independently. It also lays out what to do when money has already moved: stop communicating with the scammer, call your bank and report the incident. Sharing what happened can help protect someone else from the same approach.
Google's threat investigators have observed attackers using AI agents to automate parts of real intrusions, including searching for vulnerable systems and adjusting attacks based on the results. They also found tools that compare software before and after a security fix to work out how to attack systems still running the older version. For anyone responsible for a small business's technology, that makes lingering updates a harder risk to justify: publishing a repair can also give attackers clues about the weakness.
Mistyping an AI service's address can take you to a page that asks you to install software or complete a fake security check. Island researchers mapped 186 lookalike domains across 27 AI brands and traced several routes to malware or browser hijacking. Here, familiar AI names supply the bait. Bookmark services after verifying their addresses, and leave any supposed verification page that asks you to paste commands into your computer or install an unexpected tool.
It took five minutes and a few pounds to create a copy of Surrey Police and Crime Commissioner Lisa Townsend's voice, according to her office. In the demonstration, the cloned voice arrives on a call that also appears to come from her own number. That combination makes a family emergency scam easier to believe. If someone you recognize unexpectedly asks for money, hang up and reach them through a separate, established way of contacting them.
The people sending fraudulent emails can now buy software designed to refine their messages automatically. Varonis describes TheMailer, a phishing kit that connects to commercial AI models, creates different versions for recipients and rewrites messages after checking whether test emails land in spam. It can also turn genuine brand emails into templates. A well-written message that reaches your inbox still needs scrutiny, particularly when it asks you to sign in, approve a payment or resolve an urgent account problem.
Giving an AI assistant broad access to your Mac can expose files and let it operate other apps. Apple's macOS 27.3 beta adds more specific permission choices, clearer explanations of the access being granted and an option to require approval each time. The changes apply to new permission requests; access you previously granted stays in place. If you've been experimenting with assistants, reviewing their existing permissions is a useful job that a future update won't do for you.
For readers building apps with AI, a downloaded project's instructions deserve the same suspicion as its software. Bitdefender found 22 malicious GitHub links containing files designed to make coding assistants run concealed malware. Some configurations could also launch it automatically when a project was trusted. The malware can steal credentials and alter other projects to spread. Be especially wary of setup instructions that demand silent execution or tell the assistant not to inspect what it is about to run.
Nearly 150,000 people are being notified after attackers accessed Uber-related information held by AI infrastructure provider Fireworks, according to breach reporting. Exposed records included driver's-license details. Fireworks says stolen credentials and a vulnerability were involved in one of two incidents it has since contained. The case illustrates how personal information can reach technology providers you never chose yourself. Affected people should check their individual notices for the information involved and the monitoring and assistance Uber is offering.