Last weekend I got to experience the weird November 2026 outage-cluster firsthand, as probably quite a lot of you did. What was supposed to be a seven- or eight-hour flight trip, with one change of planes in the middle, turned into 30+ hour ordeal of iteratively canceled and rescheduled flights. I spent far too many precious hours of my life waiting in line after line at the Atlanta airport where I was trying to change planes, largely at various Delta customer service counters repeatedly rebooking new flights after the previous ones disappeared… dealing with overstressed human agents with varying degrees of helpfulness, all struggling with typing confusing codes into obstreperous semi-obsolete computer systems…
This absurd and infurating travel madness was the result of the failure of the FAA’s computer network last Friday, which shut down every departure in the US for four hours and twenty-one minutes. This is something that happens very, very rarely—it was only the second nationwide ground stop since 9/11, and a more serious one than the first. And within the same week Microsoft’s online office tools went down for most of a day, Chase Bank transactions slowed badly, Apple’s sign-in and App Store went down, and there was a bunch of smaller stuff: United’s departures stalling for a while, a Workday data center knocked out by a maintenance check, Amazon Web Services sign-ins failing in Virginia, and apparently even one of the NSA’s internal classified networks. So, like a lot of other people, I found myself wondering: why the hell did all these things crash at the same time?
I have spent a fair bit of effort since then trying to answer that—or more precisely, I have not really spent any significant time on it but have asked some of my AI agents to try to figure it out, mostly by having a few LLMs dig through the public record and test a few dozen candidate explanations. Unfortunately, I have to admit this effort to understand WTF was going on has pretty much failed. What I can offer is merely my fairly unsatisfying take so far:
- The week was very unusual, but not unbelievably so.
- There is no simple hidden common cause among all the outages that I have been able to find so far...
- Some version of the official explanation does basically kind of make sense, somewhat to my surprise.
- And yet the whole thing doesn’t sit right with me… I’m not yet fully convinced…
I’ll walk through all of that below—and then after disbursing my unsatisfying findings, I’ll ask whether anyone reading this has a deeper insight, or whether they or their AIs has figured out something I’ve missed.
How unlikely was this?
The first question I asked my agents is how unlikely it is to have this many bad things happen in so short a time. To get a baseline, I had the LLMs take ten random six-day windows from the last couple of years and count comparable outages under a rule fixed in advance. The average across this inadequate but nontrivial sample is about 0.7 every six days, or roughly forty a year, which is more than I would have thought. And last week’s seven I would round down to four or five: the NSA thing was one internal classified network among many; nobody would have cared about the Workday data center either; and the United outage lasted about thirty minutes by United’s own account. So let’s say four or five big deals in six days, two of which were huge—the FAA and Microsoft.
Run the numbers and some very crude calculations suggest a week like that should come along somewhere between once every fifteen years and once every hundred, depending on whether you count four or five. Most of that improbability is in the two huge failures landing a day apart; once you have those, the rest of the week is only about a one-in-four event. So: unusual, very unusual, but not unbelievably unusual. If there had been twenty outages that week it would be another matter.

The standard explanation
My first thought, like everybody else’s, was that it must be hacking. We are after all in what seems to be the early stages of a cybersecurity quasi-apocalypse (which I hope to help solve using my own team’s agentic neural-symbolic AI and formal verification tools, but that’s another story). But everyone involved has denied that—the companies, the intelligence agencies, the administration—and it doesn’t seem sooooo likely that all of them would lie; organizations these days are often fairly open about it when they’ve been hacked.
The official story is that these were unrelated accidents that happened to land in the same week, caused by software maintenance and upgrades. And then there is a juicier version, in which it’s indirectly the fault of AI—not that anyone used AI to hack anything, but that AI is now finding security bugs by the thousand, and the patching of all those bugs is making networks and systems fall over. Bank executives and the NSA have reportedly told Congress as much, and the Director of National Intelligence has said outright that the NSA’s network failed after somebody applied a patch for a bug in PHP, the web programming language.
This is all vaguely plausible, but unsatisfying. My first instinct was that there must be some common cause. Remember July 2024, when Windows machines all over the world crashed at once and it looked at first like some enormous attack? That turned out to have one cause: CrowdStrike, a security tool that a huge number of companies were running, had pushed out a faulty update. So I wondered whether there was a CrowdStrike behind these events too. The answer seems to be that there is nothing that simple—or at least the LLMs looked long and hard and didn’t find it.
And the narrative has some mid-sized holes in it. If it were just a matter of panicked patching, you would expect a wave of outages right after each big batch of security patches, and that hasn’t been the historical pattern. October had more security fixes than November—Microsoft’s monthly release alone went from about a hundred fixes in January to about 1,400 in October—plus an emergency over PHP, and it didn’t have this crazy cluster of outages. And the two biggest failures were not rushed patches at all, by their owners’ own accounts: Microsoft’s came from a fix it had rolled out back in July, and the FAA’s, on its preliminary account, from a contractor installing a new router.
Still, thinking it through, I have come to the reluctant and very tentative conclusion that some version of the official story might actually be right. I’ll take a few more paragraphs to walk it through a bit…
It does seem to be true that fixing bugs puts new bugs into software. Long-running industry data show about 7 percent of bug repairs introducing a new bug; a study of fixes to released operating systems found 15 to 24 percent of them were themselves buggy; and security fixes are harder than most—a small study of the Linux kernel found that incomplete ones mostly took two to seven months to be caught. So if AI is finding ten times as many bugs, then roughly speaking ten times as many bad fixes are being created, and they sit there for months before anybody notices.
The Microsoft crash is a pretty clear case. The bug that took down its gateway servers was itself created by a fix for something else, back in July, and then sat there for about four months, until a routine failover shifted extra traffic onto gateways that were already short of capacity and most of the servers stopped responding. And testing hadn’t caught it, because the tests had mostly been run on older machines with fewer cores, while the bug was at its worst on newer machines with more of them. So things are complicated: fixing a bug caused a bug, and a great tech company tested for it, and tested in a biased way that happened to miss it.
Another thing to keep in mind is the seasonality. Companies stop changing their production systems toward the end of the year—there is a code freeze over the holidays—so the week or two before Thanksgiving is when everybody pushes through their last changes. You see a lot of that here: the FAA was installing routers, Workday was doing a maintenance check, Chase and Apple were updating software. People were rushing to get things done at the last minute, and procrastination leads to errors. It shows up in the historical record too: in a public log of major cloud outages, ten of the 44 from 2020 through 2025 fell in the six weeks before Thanksgiving, twice what chance would give.
It’s also interesting to look at where things broke. Mostly it was the network plumbing—gateways, routers and sign-in systems, the channels that everything else has to pass through. That layer is being attacked, patched and replaced more this year than any other; somebody has been remotely crashing Cisco VPN firewalls since August. I’m not exactly sure why—my guess is that the newest AI models are much better than their predecessors at finding holes in this kind of software. The network-monitoring data point the same way, with outages in that layer running at about 2.6 times their spring level.
So the picture I end up with is that the risk of a big outage behaves roughly like the stock of hidden bugs multiplied by the number of maneuvers and load spikes that can set them off, and that in mid-November 2026 both numbers were at record highs. Add some bad luck—I found nothing that connects Microsoft’s failure to the FAA’s—and you get last week.
But still—if the real story is hidden bugs left behind by a year of partly-crappy patching, why would so many of them go off in the same week?

Explanations I would find more satisfying, but have zero evidence for
So what else could it be? Here are some possibilities I would find more mentally and emotionally satisfying, none of which I have any real evidence for.
It could be some adversary poking at US infrastructure, just to see how hard it is to break — maybe as a sort of prelude to something bigger. Everyone says it wasn’t an attack. But somebody has been crashing those Cisco firewalls, the web-testing firm Optimizely just disclosed a real breach—and a ransomware group has claimed an attack on JPMorgan, though the bank’s position is that Monday’s trouble had nothing to do with an attack. Etc. etc. My gut somewhat gravitates toward this explanation: some rogue agency or agent just poking around. But then why would they poke now rather than at some other time? It’s an incomplete thriller plot And there is no reported evidence of it.
It could be a single piece of software underneath everything, CrowdStrike-style. We checked the obvious candidates—a firewall hotfix, the big identity providers, a certificate authority being removed from Windows, the November Windows update itself—and there doesn’t seem to be anything there. But big companies don’t tell you everything that goes on inside them.
It could be AI coding agents, simply putting more bugs into these systems than there used to be. One large industry dataset shows incidents per code change roughly tripling in the teams that lean on AI most. But no one has said this was the cause, which to me makes it feel somewhat unlikely, because I don’t see that these companies have any big reason to deny this sort of thing happening. If anything it sounds better to say “there was this practice, which we can now change” than to say “we have no idea what the fuck is going on”. And Microsoft, for one, has stated flatly that its outage had nothing to do with AI agents.
And I wonder about the internet’s cryptographic plumbing, which a lot of people are in the middle of rebuilding. Microsoft’s cloud edge recently switched to heavier “post-quantum” encryption by default, and around December 1 the last of an older kind of security certificate, the kind used for machine-to-machine sign-in, expires. Either could cause sign-in failures at a lot of companies at once, but as far as I can see there’s no evidence for it in any of these incidents.
So, does anyone have further evidence or ideas?
A few things should help add more nuance to the story. Microsoft’s final report is due on Friday; my agents will be curious to know what that July fix was for, and why those gateways were short of capacity the day before the crash—which happened to be the day after its monthly patch release. The FAA still hasn’t settled on a root cause, and Chase, Apple and United haven’t said what exactly they were changing. And if the hidden-bugs story is right, we haven’t seen the last of this: you would expect more and more of these outages, especially with the heavier and heavier load that comes after Thanksgiving and on toward Christmas, and maybe a rush of trouble around a coming December 1 certificate deadline.
So yeah, that all just leaves me with a quite inconclusive non-conclusion. It was an unlikely combination of events… but not quite sooooo unlikely that it couldn’t be chance, plus the season, plus a year’s worth of AI-driven patching—and that might be all there is to it. On the other hand, deep down I don’t much want to accept that as an explanation. It makes a modest amount of sense, but not quite enough sense to feel compelling, and I never did find the hidden common component. So I remain somewhat confused.
I’ve also been around long enough to be able to accept that life is just going to be confusing sometimes, and you can’t really hope to resolve it all…
But still—if you know something about what happened inside any of these organizations, or you’ve spotted a useful pattern I’ve missed, my agents and I would love to hear it.