Five days in November: what failed and who could be behind it

Between 12 and 16 November the FAA held every departure in America, Microsoft 365 went down for a day, customers of five banks reported trouble and 178 organisations served a card skimmer. Each failure, each theory and what is on the record.

securityaioutagessupply-chain

Flights, Microsoft email, five banks, United and the websites of 178 organisations failed in America between 12 and 16 November. One owner calls its failure an attack.

I can't tell you one cause links the five, and no investigator has said so. Three weeks before its own outage, Microsoft reported that a China-linked group had used AI agents to try the logins of more than 2,250 of its customers. Attackers work faster than they did a year ago. I'm moving my logins to passkeys, installing patches the day they ship, auditing the outside scripts on my sites, and planning for a day with no email, bank or flights.

4 h 39

the FAA held departures across the United States, 3:36 pm to 8:15 pm Eastern on Friday

The Next Web, 16 Nov

24 h

Outlook, Teams and Copilot down, Thursday into Friday

Cybersecurity News, 16 Nov

5

banks had customers reporting trouble on Monday. One bank named a cause

AP, via Outlook India

178

organisations served card-skimming code from their own sites on Monday

BleepingComputer, 16 Nov

I've opened each source I link. Two outlets disagree in places, and there I give both numbers. No investigator has said that anyone attacked the FAA, Microsoft or a bank, and I don't claim it. I wrote this on 17 November.

Four owners found no attack and one failure was an attack

Microsoft 365

Thu 12 to Fri 13 Nov. Outlook, Teams and Copilot down for about 24 hours

FAA

Fri 13 Nov. Every departure held from shortly after 3:36 pm to 8:15 pm ET

Chase

Mon 16 Nov. Payments delayed. Customers of four more banks reported trouble

United

Mon 16 Nov. Departures held for about 1 h 40 min

Microsoft and Chase say it was not an attack. The FAA and United say they have seen no sign of one.

Optimizely

Mon 16 Nov. Card-skimming code in the scripts of 178 organisations

Optimizely confirms an attack: somebody exploited a flaw in its signup flow.

Thursday 12 to Monday 16 November 2026, a column a day. Saturday and Sunday are empty on this list.

24 h

Outlook, Teams and Copilot down, from 11:20 am Eastern on Thursday 12 November

A fix Microsoft shipped in July took Outlook down for a day in November

“This was not a cyberattack or an issue caused by AI tooling.”

Charles Lamanna, executive vice president at Microsoft, on 13 November

Still openMicrosoft ruled out an attack three days before it named a cause. It has promised a final review this week.

Microsoft shipped the bug in July. It says the bug sat in a fix it wrote for its Microsoft 365 gateway servers, and that the fix ran worldwide for more than three months before it broke anything.

  1. 1. JulyMicrosoft ships a fix for its Microsoft 365 gateway servers. It carries a hidden bug and runs worldwide for more than three months without trouble.

  2. 2. Before the outageNearly 14% of the gateway's server capacity is offline for an unrelated reason.

  3. 3. 12 NovemberTeams traffic moves onto the gateway that also carries email. The load arrives at the peak of the US working day.

80%+

of the gateway’s servers had stopped answering at the lowest point, by Microsoft’s count

Microsoft's account of 16 November, drawn as one rack of 50 servers. Grey servers were offline before the outage. Orange ones stopped answering. Two kept answering. Some came back, took the queued traffic and failed again.

Two more faults had to line up. A separate problem had taken nearly 14% of the gateway's servers offline. Then Microsoft moved Teams traffic onto the gateway that carries Exchange Online, at the peak of the American working day.

“This caused requests to build up until servers ultimately stopped responding.”

Charles Lamanna, Microsoft, 16 November. Reported by ABC News

Recovery took a day because of a loop. Engineers brought a server back, the queued traffic hit it at once, and it failed again. At the lowest point more than 80% of the gateway's servers had stopped answering. Microsoft could switch Teams, OneDrive and SharePoint to other infrastructure. It couldn't switch Exchange Online, so email had to wait for the gateway.

Microsoft's own tests missed the bug. Fox Business reports that its labs didn't match the mix of traffic and hardware in production, and that a separate fault stopped its safeguards from containing the failure. A Fortune opinion piece counted 19 services down in 74 countries.

17 of 22

FAA en route centres had unstable radar and radio on Friday 13 November

One new router held departures across America for 4 hours 39 minutes

“We were unwilling to take any risk that this propagation could actually result in a center losing connectivity.”

Bryan Bedford, FAA administrator, on 14 November

Still openThe router's owner says the installer configured it. On Monday the Transportation Secretary still couldn't say why the installation set the flood off.

The FAA has stopped the whole country three times: after the 11 September attacks, in January 2023, and on Friday.

A contractor, L3Harris, is moving the FAA's network from copper lines to fibre. It installed a new Juniper router at a radio site in El Dorado, Arkansas. The router it replaced was a Cisco made in 2001, which the FAA had meant to retire in 2006. Transportation Secretary Sean Duffy said the new router then made more than 1,000 connection attempts to other routers every four to five seconds.

New Juniper router

El Dorado, Arkansas. Installed during copper-to-fibre work. Its flood began at 3:36 pm on Friday

New Juniper router

Amarillo, Texas. Installed three weeks before. It flooded the network at the same moment

1,000+

connection attempts to other routers each four to five seconds, from the Arkansas router alone

17 of 22

en route centres had unstable radar and radio

Two routers the contractor had just installed flooded the FAA's network with connection attempts. Each square on the right is one of the 22 en route centres.

Trouble started at 3:36 pm Eastern. Radar feeds and radio went unstable at 17 of the 22 centres that handle aircraft at cruising altitude, and the FAA held departures until 8:15 pm.

A second new router flooded the network at the same moment. L3Harris had installed that one in Amarillo, Texas, three weeks before. I haven't found an explanation for why a router that had run for three weeks failed in the same minute as the new one.

L3Harris did the work during the working day.

“We didn’t anticipate that this could take down the system, which is why it was done during the day.”

Sean Duffy, Transportation Secretary, on CNBC. Reported by The Next Web

Hewlett Packard Enterprise owns Juniper, and its statement points at the contractor.

“Based on what we know today, the device functioned as it was configured to by the installer.”

Bedford said the FAA had zero evidence of an AI cause, and Homeland Security Secretary Markwayne Mullin told reporters it didn't appear to be a cyber attack. The Arkansas router now sits in an L3Harris lab in Florida. FAA staff and FBI agents are there to rule out a cyberattack, and L3Harris has blocked both sites and switched off broadcasting on each new router.

The cancellation counts differ because they cover different days. The FAA's figure is 12,462; I have it as other outlets quote it. Reuters has more than 18,000, and FlightAware splits that into about 9,200 flights on Friday and more than 6,400 on Saturday. The Fortune piece says the fault reached all 22 centres, against Reuters' 17.

5

banks had customers reporting trouble on Monday 16 November. One bank named a cause

Customers of five banks reported trouble and one bank explained it

Chase saysIts spokesman Joe Evangelisti told the AP the fault is not cyber-related.

Still openChase is the one bank of the five to name a fault. I can't tell from the statements whether its bug accounts for the reports at the other four.

On Monday morning, customers of Chase, Bank of America, U.S. Bank, Wells Fargo and Truist reported outages in the same hours. Chase told the AP that a software bug in an on-premise hardware system had delayed transactions, and that the delays reached Chase transactions for customers at other banks.

Chase

A software bug in an on-premise hardware system delays transactions.

Zelle and other payments

Chase payments to and from customers of other banks arrive late. Zelle’s operator says its own network is running.

Bank of America

Customers report trouble. Says its own services had no outage.

U.S. Bank

Customers report trouble. Says its own systems weren't directly hit.

Wells Fargo

Customers report trouble. No cause given.

Truist

Customers report trouble. No cause given.

Chase is the only bank of the five that has named a fault. Its delayed transactions included ones going to customers at other banks.

That could account for much of the day. If Chase sends your money late, your own bank's app shows the delay. Bank of America and Citigroup said their own services had no outage. U.S. Bank said the fault didn't hit its own systems, and Zelle's operator said its network was running.

I haven't found a statement from Wells Fargo or Truist, or a count of affected customers from any of the five.

1 h 40

United held its departures on Monday 16 November

United kept its planes at the gate because it couldn't get their final weights

United saysIt told the AP it had no indication of a cybersecurity issue.

Still openThe AP's report gives no cause for the failure.

A system works out each airliner's final weight before it leaves the gate. On Monday United's pilots reported that theirs had stopped working, and United held departures for about an hour and 40 minutes. Flights in the air kept flying, and United Express flew as normal.

178

organisations served the attacker's code from their own sites on Monday 16 November

One script put a card skimmer on the sites of 178 organisations

“An unauthorized party exploited a flaw in the Experimentation account signup flow to add malicious JavaScript to some customer projects.”

Optimizely, on 16 November

Still openOptimizely hasn't said how the flaw worked, how long it existed or how many visitors the code reached. No source I've read names the attacker, or says the attacker used AI.

Optimizely is the one owner of the five that calls its failure an attack. It sells a testing tool, and a customer's website loads that tool as a script from Optimizely's servers. Scott Helme of Report URI checked 1,611 of those scripts and found malicious code in 188, belonging to 178 organisations.

Optimizely

Somebody uses a flaw in the account signup flow to add code to customers’ testing scripts.

  • 1,611Optimizely scripts checked
  • 188carried the code
  • 178organisations
  • CNN
  • Zoom
  • GitLab
  • Salesforce
  • HSBC Insurance
  • BBC
  • Fox News
  • KLM
  • Air France

In the visitor’s browser, the code went after

  • Card numbers, expiry dates and security codes
  • Crypto wallet addresses, swapped for the attacker's
  • Cookies, page addresses and screenshots
  • Two iOS exploit kits, Coruna and DarkSword, loaded onto the phone
Each small square is one of the 178 organisations whose Optimizely script carried the code. Counts by Scott Helme of Report URI; the named sites were confirmed by BleepingComputer.

If you opened CNN, the BBC or a Zoom page on Monday, the site gave your browser the attacker's code. The attacker wrote it to copy card numbers, expiry dates and security codes, and to swap crypto wallet addresses for their own. It could take cookies, page addresses and screenshots, and it loaded two iOS exploit kits called Coruna and DarkSword.

About $7,000 reached the attacker's wallets. Some of that arrived before this attack began, so BleepingComputer couldn't say how much the attacker stole. Optimizely says it has fixed the flaw, removed the code and reset the affected accounts. At 21:52 UTC on Monday, BleepingComputer found the code in one of the 188 files.

Seven theories and what would settle each one

I went looking for the best argument behind each theory, including the ones I expected to dismiss. The board shows where each stands on 17 November, and each row links to its section.

A state

A named person or agency has claimed it

A known mechanism fits

Someone has published evidence

An AI that no one steers

A named person or agency has claimed it

A known mechanism fits

Someone has published evidence

DARPA or a government test

A named person or agency has claimed it

A known mechanism fits

Someone has published evidence

Cover for something else

A named person or agency has claimed it

A known mechanism fits

Someone has published evidence

A competency crisis

A named person or agency has claimed it

A known mechanism fits

Someone has published evidence

No culprit: old systems and bad timing

A named person or agency has claimed it

A known mechanism fits

Someone has published evidence

A 90-day blackout from 17 November

A named person or agency has claimed it

A known mechanism fits

Someone has published evidence

My reading of the sources linked in each section, as of 17 November. A solid square means yes. For the row with no culprit, the published evidence is each owner’s own preliminary account.

A state had the access and the motive, and no agency has accused one

No accusationNo US agency, affected company or named expert has blamed a foreign government.

Best caseUS agencies say China's Volt Typhoon group has sat inside American communications, transport, energy and water networks, in some cases for five years. Three weeks before the outages, Microsoft caught a China-linked group running AI agents against its customers.

Weakest pointEach failure has a documented cause that needs no attacker. Aviation and banking systems run apart from Microsoft's cloud. No group has claimed the week, and no one has reported a ransom note, stolen data or malicious code.

Would settle itThe FBI's examination of the Arkansas router, Microsoft's final review, and a root-cause report from Chase.

The strongest version comes from a Fortune opinion piece by four authors, two of them retired colonels. They ask the question and stop short of an answer.

“Could China and other hostile foreign actors be targeting U.S. infrastructure? It’s too early to know for sure, but the country deserves far more transparent answers than we have received so far.”

Jeffrey Sonnenfeld, Steven Tian, Cedric Leighton and Vito Errico, in a Fortune commentary, 13 November

Their case rests on access. Volt Typhoon spent more than 300 days inside the Littleton Electric Light and Water Departments in Massachusetts, from February to November 2023. Europe had trouble in the same week: France 24 reports attacks on supermarket refrigeration in seven countries, with as many as 700 French supermarkets affected.

IVN makes the plainest argument on the other side.

“Banking and aviation systems run almost entirely apart from Microsoft’s consumer cloud, and a failure in one does not naturally propagate to the others.”

Ethan Handelman, Independent Voter News, 16 November

My own doubt is about cost, and it is my reasoning, with no source behind it. Volt Typhoon's access is worth most in a war. I can't see a state that hid in American networks for five years spending that access on a four-hour ground stop.

The Homeland Security Secretary floated an AI, and the FAA had expanded one three days before

One official floated itMarkwayne Mullin relayed an unnamed expert's view on Fox News on 16 November. He named no program.

Best caseOn 10 November the FAA announced that its AI routing tool, SMART, would expand to all 22 en route centres. The ground stop came three days later. Copilot was among the Microsoft products that went down.

Weakest pointThe FAA administrator says the agency has zero evidence of an AI cause. Government officials and SMART's developer, Air Space Intelligence, say the tool had no part in the outage. Microsoft says AI tooling didn't cause its failure.

Would settle itThe FAA publishing what SMART did at 3:36 pm on Friday, beside the logs from the Arkansas router.

Mullin told Sean Hannity he was 98% certain the week was no cyberattack. He then said a technology expert, whom he didn't name, had told him super intelligence could be contributing, and that this kind of made sense to him.

“We used to have backup systems and the backup systems are overwhelmed by the new systems now.”

Markwayne Mullin, Homeland Security Secretary, on Fox News, 16 November. Reported by Outlook India

Super intelligence sounds like a claim about a machine smarter than people. In this administration it is the word for AI. Three days before, Mullin had corrected a reporter who said artificial intelligence, and Raw Story reports that an executive order replaced the term across federal agencies.

“We call it SI around here, remember? SI.”

Markwayne Mullin, to reporters, 13 November. Reported by Raw Story

Read with that in mind, Mullin's argument is about load: new systems overwhelming old backups. The FAA describes the same thing, with a router where he put an AI.

Shattered set three tests for an AI headline on 15 November. Has the company said an AI system caused it, has a named official said so, and does a technical explanation work without AI? It answered no, no and yes. Mullin moved the second answer to maybe the next day, and Marjorie Taylor Greene asked the same question on X.

I looked for DARPA and a government test, and found a real programme and no claim

No claim foundI found no article, official or document that ties DARPA or a US exercise to the outages.

Best caseDARPA does build this technology. Its AI Cyber Challenge produced AI systems that find, exploit and patch software flaws at machine speed. A federal alert test and a Pentagon plan to run bases off-grid for 90 days fell in the same weeks.

Weakest pointThe challenge is a public competition on open-source code. A test that cancelled more than 18,000 flights and cost what the Transportation Secretary called billions of dollars would have left a paper trail, and no one has leaked one.

Would settle itA document, a contract or a witness. I have none of the three to weigh.

The programme is real, and the Congressional Research Service describes it.

“DARPA’s AIxCC competition is focused on developing AI systems capable of autonomously identifying, exploiting, and patching software vulnerabilities at machine speed. The goal of the two-year challenge is to harden critical infrastructure by enabling proactive, autonomous cyber defense.”

Congressional Research Service, Agentic Artificial Intelligence and Cyberattacks, updated July 2026

The test is real too. At 2:20 pm Eastern on 17 November, FEMA and the FCC test the Emergency Alert System. Snopes notes that they announced it in September and that the law requires one at least once in three years. Snopes also describes a Department of Defense plan to keep military bases running for at least 90 days without civilian infrastructure, and Greene set that plan beside the ground stop in a post on 13 November.

I ran four searches for a page that connects DARPA, a drill or a US agency to the five failures and found none. One true thing sits inside the theory, though. The government's own contractor installed the router that stopped the flights, and the White House cyber director expects more of the same as agencies replace old systems.

“No, this is not any sort of coordinated, nefarious cyberattack from a malicious actor.”

Sean Cairncross, White House cyber director, 16 November. Reported by Semafor

A former congresswoman says the chaos is cover, and offers no evidence

One claim, no evidenceMarjorie Taylor Greene, on X, 16 November.

Best caseAn outage is a good moment to move something unnoticed, and three came in five days. Greene suggested deleted records, moved aircraft and moved money.

Weakest pointIBTimes reports that she offered no evidence of deleted records, undisclosed flights or improper transfers. One insider would need the same reach at Microsoft, the FAA and Chase.

Would settle itAudit logs from the three organisations, which their own reviews should cover.

Greene called the Microsoft outage a good time to delete information, IBTimes reports, and added a line that sums up the theory.

“Chaos keeps everyone distracted.”

Marjorie Taylor Greene, on X, 16 November. Reported by IBTimes UK

A cheaper version needs no plot. Mike Chapple, a professor at Notre Dame, told USA Today that the FAA outage began with a faulty component or an installation error by a government contractor. An error by an insider fits the record. A plan by one has nothing behind it yet.

The people who could tell a bad router from a bad actor have left

Argued in opinion piecesA Fortune commentary and a Substack newsletter make the case. No official does.

Best caseCISA went from roughly 3,300 staff to about 2,400 by December 2025 and lost its red team. More than 1,500 special agents have left the FBI. The FAA fired 352 probationary employees, and their union says radar and navigation technicians were among them.

Weakest pointNo source says a departure caused an outage. A contractor installed the FAA's router, and Microsoft and Chase are private companies with their own engineers. The FAA says its cuts came to under 1% of staff and spared controllers.

Would settle itSworn testimony on how many technical staff left, what they did, and who checked the router installation.

The figures come from two opinion pieces, Fortune's and a Substack newsletter by William J Russell, so treat them as those authors' numbers. Russell's argument is about the investigation, and it is the one I keep coming back to.

“When officials say there is no evidence of an attack, the honest follow-up question is how hard the people who would find that evidence are able to look.”

William J Russell, The Political Will Report, 17 November

I think this theory has the most behind it and gets the least attention. It doesn't say who broke what. It says the country has fewer people who could find out, and that part is a headcount. Duffy has asked Congress for $50 billion for a new network, and The Next Web reports that House Democrats say his department has spent $1 billion of the $12.5 billion it got last year.

Old systems and tight coupling explain four of the five, and an upgrade caused the biggest one

The official accountMicrosoft, the FAA, Chase, United and the White House cyber director each describe a fault.

Best caseEach owner has described a fault that needs no attacker. In 2024 the GAO rated 105 of the FAA's 138 air traffic control systems unsustainable or potentially unsustainable. A corrupted database file grounded US flights in January 2023, and one bad update crashed 8.5 million Windows machines in July 2024.

Weakest pointFour national failures in five days is a lot of bad timing, and Microsoft ruled out an attack three days before it named a cause.

Would settle itMicrosoft's final review, the FBI's finding on the router, and a published cause from Chase and United.

Ali Butt, a professor at Virginia Tech, told USA Today that a routine change set each failure off.

“In tightly connected systems, a small change made without fully mapping its downstream effects can spread far beyond where it started.”

Ali Butt, professor at Virginia Tech. Reported by USA Today

The usual telling of this theory blames old equipment, and the record cuts against that. The 2001 Cisco router didn't fail. The new Juniper router that replaced it did, so the upgrade was the event. The GAO figure comes from CYBR.SEC.Media, which also asks readers to wait for evidence before naming a culprit.

A retired FBI agent put the case against coincidence better than I can.

“This is almost like the movie 'Casino' when all the sudden three slot machines hit… the same slot machine hits three times something doesn't make sense, and we need answers.”

I'm with Augenbaum on the instinct. I don't believe in five coincidences, and I can't prove a thing.

The 17 November blackout was a hoax that borrowed a real test

FalseFrance 24 and Snopes traced it. Snopes left it unrated until the day had passed.

Best caseThe date is real. FEMA and the FCC test the Emergency Alert System at 2:20 pm Eastern on 17 November, and a 2023 episode of The Simpsons does open a blackout on 17 November.

Weakest pointThe test interrupts TV and radio for about a minute. The episode's blackout comes from a storm. Crypto accounts pushed the claim and told people to move their savings into crypto.

Would settle itThe date. A blackout that starts on 17 November is easy to check on the 18th.

The rumour said a 90-day blackout would follow the alert test and take electricity, the internet and banking with it. France 24 found crypto accounts claiming the test was preparation for a cyberattack on the banks. Snopes asked the question that ends it.

“If the supposed blackout is supposed to be an act of sabotage by a foreign adversary, why would they have announced the attack beforehand?”

Jack Izzo, Snopes, 17 November

Snopes also explains the spread: after a Microsoft outage, a ground stop and a day of bank trouble, the claim felt plausible. The EU had told households to keep emergency food and water, and the UK had launched a campaign called Be Ready. A fake blackout fitted the week's real news.

AI has made attackers faster and the documents say so

On 22 October, three weeks before its own outage, Microsoft said a China-linked group it calls Storm-0940 had used AI agents against its customers.

151,000

failed sign-in attempts on 327 accounts at eight organisations, 10 to 14 October. Microsoft saw no successful login.

2,250+

Microsoft 365 tenants tried next, with stolen ConnectWise application credentials and an open-source AI agent framework, at up to 2,900 attempts a minute. Most attempts failed.

Under 3 h

from picking a target to privileged cloud access, in some cases. The group kept that access at what Microsoft calls numerous organisations.

Microsoft's account of Storm-0940, reported by The Register on 22 October. Microsoft hasn't said how many customers the group got into.

The group has operated since at least 2021. It used ARTEX, an open-source penetration-testing framework whose AI agents run the whole job, from reconnaissance to credential attacks and data theft. Its targets included government agencies, defence contractors and banks. The Register reports that the methods were old ones, stolen credentials and password spraying, and that Microsoft says AI let the group run them faster. Microsoft has seen attackers from several regions using the same framework.

Microsoft 365 then went down for a day. Microsoft says a bug did that, and I've found no source that links the two events.

An attacker has a working exploit for a critical flaw

5 days

A company has fixed that flaw on its outside-facing systems

30 to 60

Medians from Microsoft's Digital Defense Report, as Barracuda summarised it on 28 October. The bars share one scale of 60 days.

Microsoft's annual threat report puts numbers on the race. An attacker has a working exploit for a critical flaw about five days after it becomes public. A company takes 30 to 60 days to fix the same flaw on its outside-facing systems. For the fastest cases, Microsoft's median from discovery to a working exploit is under 24 hours.

141

flaws Google saw attackers exploit from January to August 2026, against 127 in the whole of 2025

Google GTIG, via CSO Online

80 days

median from a flaw's publication to confirmed exploitation in 2026, down from 120 in 2025

VulnCheck, via CSO Online

4 to 7

months that free, downloadable AI models trail the best closed ones on cyber tasks

UK AI Security Institute, 17 July

27 sec

the fastest breakout CrowdStrike recorded from a criminal intruder in 2025. Its average was 29 minutes

CrowdStrike 2026 Global Threat Report

72 h to 24 h

average time for hackers to turn a raw flaw into a working exploit, last year against 2026

Recorded Future, via Security Boulevard

Each figure measures a different thing, so I haven't combined them. The open-model gap matters most to me. The models that do this work are free to download, a few months behind the best, and no vendor can switch them off.

The first documented case came in 2025. The Congressional Research Service records that Anthropic caught a Chinese state-sponsored group using its Claude tools to automate 80% to 90% of an espionage campaign against about 30 organisations. The same report notes that some researchers doubt the campaign was as autonomous as Anthropic described.

Last month Anthropic's red team showed how an AI might one day spread through cheap gadgets by itself. Their demo ran a model across 15 off-brand TV boxes that cost about $55 each. They say small models may be months away from doing the whole job, and they are plain about where things stand today.

“We don’t yet have a full rogue replicator, but agents are becoming increasingly capable of completing the individual steps in the replication chain (and passing on their adaptations to future copies of themselves).”

William Zhou and Daniel Freeman, Anthropic Frontier Red Team, 12 October 2026

No source says the attacker at Optimizely used AI, and I don't claim it.

What the week tells us and what I'd change this week

Each of the five failures began with something routine. Microsoft shipped a fix, a contractor swapped a router, and somebody filled in a signup form at Optimizely. A small change in one place reached the whole country because so much of the country runs through so few systems.

“When one piece fails, it fails at national scale.”

Patrick Moorhead, founder of Moor Insights and Strategy. Reported by USA Today

From outside, you can't tell an attack from an accident. Both look like a spinning wheel on your banking app, and both come with a statement that says it wasn't hackers. Microsoft made that statement three days before it could say what had broken. My bet is that the official accounts hold for most of this week. I'd trust them more if CISA, the agency that checks such accounts, hadn't lost about a third of its staff, by Fortune's count.

More of this is coming, and the government says so. The White House cyber director told Semafor he expects more disruptions as agencies replace old systems. The experts USA Today interviewed expect failures to come more often and hit harder.

Attackers are faster than they were a year ago, and Optimizely shows who pays for that. The people whose cards it copied opened a news site. I'm changing four things on my own systems this week.

Move your logins to passkeys

Storm-0940's agents tried sign-ins at up to 2,900 a minute. A passkey gives them no password to guess or phish.

  • Start with email, your bank and your domain registrar.
  • Remove text-message codes as a fallback where the service allows it.

Patch the day the fix ships

Microsoft's median attacker has an exploit in five days. Its median company takes 30 to 60 days to fix the flaw.

  • Turn on automatic updates for phones, browsers and your home router.
  • Give servers a 48-hour limit for critical fixes on anything that faces the internet.

Treat outside scripts as your own code

178 organisations served a card skimmer because one vendor's signup flow had a flaw.

  • List each third-party script your site loads and remove the ones you can't name a use for.
  • Host the rest yourself or pin them with Subresource Integrity, and set a Content Security Policy.

Change things in stages and rehearse the bad day

The FAA's contractor swapped the router during the working day, and Microsoft's fix met production traffic its labs hadn't seen.

  • Roll each change out to one site or 1% of users first, with a way back.
  • Run a drill for a day with no email, bank app or flights, and write down what you lacked: a second email provider, some cash, offline copies of bookings.

One more habit costs nothing. Scammers follow outages, and a retired FBI agent gave WKRN the rule I'd use.

“Be on the lookout for anyone sending emails trying to take advantage of the situation. Stop. Pause. Verify. And confirm.”

Three reports should answer most of what's open. Microsoft has promised its final review this week. The FBI is in a lab in Florida with the Arkansas router. Maxine Waters, the top Democrat on the House Financial Services Committee, has asked the banks to say what went wrong. I'll add each one here when it lands.

Something strange happened in those five days. I'd make the same four changes whether the cause turns out to be a state, a bug or a router.

Building something like this?

30 minutes, no pitch.

More writing