At 3:54 pm Eastern on 13 November, the Federal Aviation Administration stopped every departure in the United States. Communications at several of the en-route centers that handle aircraft between airports had been disrupted. Nationwide ground stops are rare; earlier ones followed 9/11 and a system failure in January 2023. More than 18,000 flights had been canceled by 16 November, according to FlightAware figures reported by Fox Business.
Officials have since given a first account. A contractor, L3Harris, is replacing the copper lines in the FAA's telecommunications network with fiber. Officials said routers replaced during that work in El Dorado, Arkansas, and Amarillo, Texas, are under investigation, and that the Texas router began flooding the network and overloaded routers at other centers. Both routers are being tested in a lab. The FAA has disabled the function involved on those routers and taken similar precautions on about 60 others. Whether the cause was the installation, the equipment or something else is not yet known.
On 12 November, Microsoft 365 failed around the world for about a day. Microsoft's preliminary report traces it to a fix rolled out in July. The fix went through staged deployment and ran without trouble for three months, but it carried a latent defect that was worse on newer hardware under heavy load, and Microsoft's load tests, run on weaker hardware, missed it. Microsoft says the outage was not related to cyber activity or AI agents.
On 16 November, outage trackers lit up for Chase, Bank of America, Wells Fargo, U.S. Bank and others, and Chase customers reported that they could not move money. Chase blamed a software bug in its own on-premises hardware systems and said there was no evidence it was cyber-related. Chase said other banks may have been affected where they were party to a transaction with it, but the problem was not systemic. An FBI official, quoted by Fox Business, traced the outage to a software release Chase rolled out that day.
In September I wrote here that AI raises how much an institution produces without raising the hours available to check it. The first thing to saturate, I argued, is not the model but the institution's own capacity to verify what the model does. None of these investigations is finished, and I have no private data on any of these events: everything here comes from public statements, news reporting and public records. So none of it proves anything yet. But two of the failures fit the pattern that argument predicts, and so does a decision under way at Microsoft that I think is a mistake.
Is this an attack?
Large failures in aviation, cloud software and banking within five days invite a single explanation, and many people online have offered one: a coordinated attack. The Homeland Security Secretary says he is nearly certain it is not one and that the incidents are unrelated. On 17 November the Director of National Intelligence, Jay Clayton, said on Fox Business, "We do not see the indicia of a coordinated cyber attack." He also saw no sign of state-sponsored activity (Table 1).
Table 1. What each company and investigator says, as of 17 November 2026
| Date | What failed | Cause, as stated so far | Said by | An attack? |
|---|---|---|---|---|
| 12 to 13 Nov | Microsoft 365, worldwide | A latent defect in a fix rolled out in July, missed by load tests run on weaker hardware | Microsoft, preliminary report | Not related to cyber activity, Microsoft says |
| 13 Nov | FAA en-route communications, then a national ground stop | A router replaced during contractor fiber work began flooding the network; both routers involved are under lab test; cause not yet known | Transportation Secretary and FAA Administrator | No indication of a hack; not ruled out |
| 16 Nov | Chase payments; other banks possibly affected, Chase says | A software bug in on-premises hardware systems; an FBI official, quoted by Fox Business, ties it to a release that day | Chase; FBI official | No evidence it was cyber-related, Chase says |
| 16 Nov | Scripts that Optimizely serves to at least 178 organizations' websites | An attacker exploited a flaw in Optimizely's sign-up process and wrote malware into the scripts | Optimizely; Report URI researchers | Yes, Optimizely says |
| 17 Nov | A classified NSA network | Patching after a software vulnerability was identified; some systems did not come back up as expected | Director of National Intelligence, on Fox Business | No sign of a coordinated or state-sponsored attack, he says |
If one shared supplier sat behind the FAA, Microsoft and Chase failures, I would expect the causes to have something in common. Microsoft and Chase each name a different mechanism, the FAA's account so far involves different technology again, and no common vendor has been named. One attacker could hide better than that, which is why the operators' own findings matter more than the pattern. Transportation Secretary Sean Duffy says there is no indication of a hack but has not ruled one out, and the FBI is helping the investigation. What would change my mind is specific: a credible claim of responsibility, the same indicator found at more than one victim, a finding that someone changed the routers without authority, or a response that looks like containing an intruder.
On 16 November, researchers showed what an attack looks like when it is one. Scott Helme's team at Report URI found malware in scripts that Optimizely, a website-testing company, serves to its customers; at least 178 organizations' sites were serving it. Within hours the researchers had published the attacker's domains so defenders could block them. That is the same indicator at many victims, found fast. Nothing like it has surfaced for the outages so far. It is also the reach problem from the other side: one compromise at one supplier reached at least 178.
The same week brought a reminder that AI can multiply output for attackers too. An FBI flash advisory, reported by TechCrunch on 16 November, warns that hackers have used Artex, an open-source tool that uses AI to automate scanning a network, exploiting simple weaknesses such as stolen passwords and moving toward sensitive data, against financial, government, technology, defense, telecommunications and legal targets. In some cases, TechCrunch reports, attacks went from targeting a system to compromising it in under three hours. Barracuda, a security company, says an intruder whose methods match Artex spent 44 hours inside about ten of its cloud accounts on 1 and 2 November, using an access key it had given an outside AI vendor and never retired. No one, including the FBI, has tied Artex to the failures in Table 1.
A cluster on its own is weak evidence of a common cause. With thousands of large operators changing their systems every day, some week can be expected to produce one like this by chance. For the first hours, an attacker and a bad change look alike from inside. When Cloudflare took down large parts of the internet in November 2025, its own engineers first thought they were under a huge denial-of-service attack. The cause was a database permissions change that doubled the size of a configuration file. Either way, the protection is limiting how far one device or one change can reach.
One change can reach every site at once
The FAA account is the clearest case so far. Replacing a copper line is routine work, and more than 3,500 have been done. But on officials' account, a change at one or two remote sites reached most of the country's en-route centers. When each change carries a small risk and the rare failure a national cost, the effort spent checking a change should scale with how far it can reach, not with how often the same work has gone fine before. The FAA's response points the same way: it has treated dozens of other routers as at risk. Centralized platforms tend to make every change reach further, and I see nothing to suggest the hours available to review each one grow to match.
Microsoft's account shows the same thing at a different layer. Its load tests ran on weaker hardware than production, so they missed the defect. And when the safeguards meant to contain a fault failed for a separate reason, the systems built to route traffic around a failing server carried one fault across the fleet, until more than 80 percent of servers were unresponsive. Checking a change means checking it under the conditions it will meet.
The ground stop was also the system doing its job. When communications became unreliable, the FAA stopped traffic. That cost days of disruption, and no accident has been reported. My original piece was about the opposite failure, the quiet one, where a saturated number flattens and reads as calm.
A broken channel can read as catastrophe
The mirror image showed up too. On the evening of the ground stop, posts claimed a nationwide weather radar outage. When I checked the Weather Service's public data feed at about 5:40 pm Eastern, the radar network was up and a station's data was current, although the public website was failing and the Weather Service reported communications problems nationwide.
On 16 November, outage trackers showed spikes at six large banks at once, which looks like a failing banking system. But those trackers count complaints, not failures. They cannot tell "my bank is down" from "my payment to a Chase customer is stuck." Wells Fargo did post its own notice, without a cause. One origin with spillover and six separate failures look the same on a tracker, and that is the point.
In September I described what I call the ascertainment ceiling: a saturated reporting channel reads as calm. A channel that counts alarm rather than failure can read as catastrophe just as easily.
Rolled out before the evaluation finished
In September the FAA began using SMART, an AI tool from Air Space Intelligence that recommends reroutes, around the Washington airports. On 9 October more than two dozen lawmakers asked the FAA to pause it and explain its planned 90-day evaluation. By early November it covered the Jacksonville, Miami and Atlanta centers. On 10 November, 50 days into the 90, if the evaluation began at launch, the FAA announced that SMART would cover all 22 en-route centers.
FAA · SMART rerouting tool · 2026
The national rollout was announced on day 50 of a 90-day evaluation, counting from launch.
- 121 Sep. SMART starts around the Washington airports.
- 29 Oct. More than two dozen lawmakers ask the FAA to pause it.
- 310 Nov, day 50 from launch. The FAA says SMART will cover all 22 en-route centers.
- 4About 20 Dec. The 90-day evaluation ends.
- By early November (dates not given): the Jacksonville, Miami and Atlanta centers.
The timing invites a conclusion the evidence does not support. The developer says SMART did not cause the ground stop, and nothing in the officials' account involves it. What failed was the communications network; SMART plans traffic flow.
The FAA describes specialists reviewing the tool's recommendations, with local managers deciding whether to act on them. I would expect that to be harder to sustain across 22 centers than across a handful of facilities. The question I would want answered is the second one from my original piece. Of everything the system recommended in October, what share got a real second look, and has that share risen or fallen since 10 November?
More to check
The best large-scale evidence I know of is Google's DORA research on software delivery. Its 2025 survey of nearly 5,000 technology professionals found that 90 percent use AI at work and 30 percent have little or no trust in the code it generates. AI adoption is associated with higher delivery throughput and, still, with lower delivery stability: teams that use more AI tend to ship faster and break more, which is what you would expect if output scaled and checking did not. It is a survey, so it shows association rather than cause.
Public data shows the number of flaws to fix rising outside any one company, whatever the cause. Publicly disclosed software flaws are normally given a public record, a CVE, and the US National Vulnerability Database published about 138 a day in January 2025, about the same in January 2026, and 550 a day in October 2026, four times the rate of October 2025. The CVE Program's own quarterly counts show the same rise. Each record is something an operator has to assess, and many need a patch that has to be tested and rolled out without breaking anything else. The rise is spread across hundreds of publishers, including Microsoft, Oracle and GitHub (Table 2), so it is not one publisher changing how it reports. A researcher at the security firm Tenable says that for most of 2026 AI "has been building bigger haystacks" in Microsoft's monthly patch releases. AI is likely part of the rise, but the public counts cannot say how much.
Table 2. Vulnerability records published in October, by publisher
| Publisher | October 2025 | October 2026 |
|---|---|---|
| VulnCheck (publishes records for other vendors' products) | 157 | 2,473 |
| GitHub security advisories | 272 | 1,537 |
| Microsoft | 180 | 1,475 |
| Oracle (its quarterly update falls in October both years) | 78 | 1,078 |
| All publishers | 4,273 | 17,051 |
Source: NIST National Vulnerability Database, by the organization that published each record, rejected records excluded. Single months are noisy for vendors that release on their own cycles.
The list of flaws known to be exploited has not risen to match. In the first ten months of 2026, published vulnerability records more than doubled while CISA's additions to its known-exploited list rose by about a third. That list counts flaws, not attacks, and CISA decides what goes on it, so it cannot rule out more attacks. But it gives no sign of a fourfold surge. What the data shows is many more flaws to assess and fix. That is the treadmill: more changes to check, arriving faster, and no public measure that the capacity to check them has grown to match.
Public records · January 2025 to mid-November 2026
Many more flaws to fix. No matching rise in flaws known to be exploited.
New vulnerability records published per day, monthly average
Added to CISA's known-exploited list, per month
The last month is partial · rejected records excluded
The closest public account of the treadmill causing outages came on 17 November. The Director of National Intelligence tied an outage at the NSA to patching after a software flaw was identified, and said "lots of institutions" are patching because vulnerabilities have been identified. He did not say AI found that flaw. Later on 17 November, The Washington Sun reported, citing three people familiar with the conversations, that Wall Street executives and the NSA had told members of Congress that the week's outages were caused by attempts to fix vulnerabilities identified by AI. The report does not say which outages were caused this way: it mentions Chase and Microsoft only for their own explanations, and says it is unclear what caused the FAA failure. Microsoft traces its failure to a fix made in July, neither company has said its failure came from a security patch, and nobody has confirmed the report on the record. If it holds, some of the week's outages shared a cause after all, though not an attacker, and it is the treadmill in one line: AI, in the report's words, "not causing the disruptions but identifying gaps that are difficult to fix."
I feel a small version of this in my own work. The outbreak tool Apolitical featured in August was built by one person using AI coding agents, and checking what the agents produce takes more of my time than producing it.
A buyout does not choose who leaves
Microsoft forecasts 175 billion dollars of capital spending in 2026 and, like other large technology companies, is spending heavily on AI data centers and chips. In July it cut about 4,800 roles and said those roles "are not being replaced by AI." On 26 October it offered voluntary buyouts to about 60,000 people, roughly a quarter of the company, with no target; according to Reuters, the offer is partly meant to help fund its AI investment. Eligible groups reportedly include the Windows + Devices division. Employees decide in early December and leave on 31 December.
On the DORA evidence, teams that use AI tend to deliver more. What I do not see speeding up to match is the work of knowing which changes are dangerous: review, testing and staged rollout, done by people who have watched things break before.
A voluntary program does not pick who goes. In a broad buyout, I would expect the people with the best options elsewhere to be among the first to leave, and in engineering those are often the most experienced. They are usually also the reviewers. Nobody outside Microsoft knows who will accept, and I am not claiming to. But if these programs take the people who know which change is dangerous, the company is funding its AI build-out with the capacity that, on the evidence so far, has not kept pace. I think that would be the wrong saving.
According to the Reuters source who described the eligible groups, security staff were kept out of the October offer. In 2024 a federal review board had found Microsoft's security culture "inadequate." The company can ring-fence a checking function when it decides to. The open question is whether reliability, testing and code review get the same protection.
Nothing I have seen connects any of this to Microsoft's 12 November outage. The October leavers have not gone yet, and nothing ties the earlier departures to the change Microsoft blamed.
There are fair objections. The cuts may miss the checking queue: July's fell mostly on Xbox, sales and management layers, and some engineering teams reportedly declined the buyout. AI checks as well as produces: a researcher using AI found one of the exploited flaws Microsoft patched in November. And the capital pressure is real. My claim holds only if checking capacity grows more slowly than output, and the DORA stability finding is the best evidence I have that, for now, it does. The argument is not that companies should never cut, but that they should not cut the checking queue, which I think is already saturating, and should be able to show, with a number, that they have not.
Public institutions face the same trade
Public institutions face the same decision. AI can make a team more productive. If the gain is booked as fewer staff, I would expect the people who leave not to be chosen by what they do. The budget would improve straight away. Any shortfall in checking would appear later, as incidents, and be easy to attribute to something else.
That is the calm number at the level of staffing. The assurance figure can stay green because it counts what was processed, not what was examined.
The questions I would ask:
When you booked the productivity gain from AI as fewer staff, did anyone check whether the people leaving were the ones checking the work?
What was your second-look rate before the last restructuring, and what is it now?
How long was the evaluation window for your last AI deployment, and did the rollout wait for it?
When a contractor changes something on a network every one of your sites depends on, who on your side decides when it happens, who can say no, and how quickly could you tell that change from an attack?
AI was used to research and draft this post.