KKReckons Special Report · Cyber & resilience

Breaking, not ending: the week the systems failed

Email went dark, planes stood still, banks stalled, and the internet decided the world would end at 2:20 on Tuesday afternoon. It didn't. But behind the noise, something real is speeding up.

Six days
Thursday, 4:17 pm in London

Outlook stops loading. Within two hours, more than four in five of the servers in front of Microsoft's email have frozen.

Friday, 3:30 pm in Washington

Radar feeds and radio links flicker at 17 of America's 22 en-route control centres. Every departure in the country is told to stay on the ground.

Monday, 4:21 am UTC

A few lines of malicious code slip into a script on the BBC's website. Over the day, the same code turns up on Zoom, CNN, HSBC and 175 other organisations' sites.

Tuesday, 2:20 pm Eastern

A routine one-minute alert test, scheduled since September, which millions had been told online was the start of a 90-day blackout.

This report sorts what broke from what was attacked, and what was just rumour. Then it looks at the real threat underneath, and what experts say would contain it.

The short answer

No single cause. Two real trends colliding.

  • The big outages were self-inflicted. Microsoft, the FAA, Chase and Apple each traced their failure to their own software, update or equipment work. The NSA hasn't given a cause, but people familiar with it say there's no sign of an attack.
  • The attacks were real but separate. Japan's Fugaku supercomputer was breached, and a hijacked web-testing service planted card-stealing code on 178 organisations' sites. Nothing links either to the outages.
  • The doomsday was a hoax. The "17 November blackout" grew out of a routine emergency-alert test.
  • The state attacks are in Europe, and in the freezer aisle. Hackers heated supermarket fridges in about 700 French stores in October and dozens more across Europe on 2 November. France blames an Iranian Revolutionary Guards cyber unit. Governments from London to Berlin are telling households to keep days of supplies.
  • The worrying part is speed. The FBI says hackers now use an AI tool, ARTEX, to go from first target to admin access in under three hours. An AI agent found a critical Java flaw in about an hour. Leaked cloud keys get used in seconds.
~24 hExchange email disrupted worldwide
$7 bnthink-tank estimate of lost productivity from the Microsoft outage
9,000+US flights cancelled on Friday alone
178organisations whose sites served stolen-card code
7.86 sfastest time from a key leaking online to its first use
0links found between the outages and any attack
1 · The week, verified

What broke, who it hit, and what the owners say

Each entry is checked against the company, agency or researcher involved. Times are as reported; status is as of Tuesday night, Pacific time.

THU 12 NOVMicrosoft 365 downOutlook, Teams, CopilotEmail out ~24 hours;80%+ gateway servers downFRI 13 NOVFAA ground stop9,000+ flights cancelledFugaku intruderdetected; logins triedSAT 14 NOVFugaku shut downresearcher keys stolenMicrosoft restoredafter ~24 h for emailSUN 15 NOVAirlines catch up2,000+ more cancellationsBlackout rumoursspread on TikTok, XMON 16 NOVChase down ~8 hsoftware bugSkimmer on 178 sitesvia OptimizelyTUE 17 NOVNSA Net, Apple downno sign of attackJava4Shell; alert testAI-found bug; hoax dateBroke itselfAttackAI flaw or rumour
KKReckons graphic. The six days at a glance; details for each incident follow.
Nov 12–14
Microsoft 365: Outlook, Teams, Copilot down worldwide
Exchange Online was disrupted for almost 24 hours; the last customers came back on Saturday. Microsoft blames three things together: a bug hidden in a July fix, extra load, and capacity already out of service (see section 2). No sign of malicious activity.
Who felt it: about 430 million users in 70+ countries. NHS staff in England struggled to pass on patient details from 111 and 999 calls; Brazil's Supreme Court paused legal deadlines; some Australian schools couldn't run online classes.
Own software
Nov 13
FAA grounds every departing US flight for ~4½ hours
The FAA says contractor L3Harris was installing a new router at a remote site during a routine copper-to-fibre swap. Bad data spread across the air traffic network until the sites were blocked. Officials ruled out a cyberattack; the routers are being examined with FBI agents present.
Who felt it: FlightAware counted more than 9,000 cancellations and nearly 25,000 delays on Friday, and over 4,000 more cancellations by Saturday morning. Travellers slept on terminal floors from Miami to Chicago.
Own upgrade (contractor)
Nov 10–17
Japan's Fugaku supercomputer breached and shut down
An intruder got in through Fugaku's public status dashboard, moved across RIKEN's network and installed malware. RIKEN suspended Fugaku on 14 Nov and disclosed on 17 Nov. Attacker unknown.
Who felt it: up to 4,597 researchers, whose names, emails and IDs were taken or may have been, along with SSH keys they kept for other supercomputers. The attacker has already tried to use some of those keys.
Confirmed attack
Nov 16
Chase online banking down for nearly eight hours
Chase blamed a software bug and said it was not cyber-related. Customers of other big banks reported problems the same day. United Airlines also had a departure-system fault; no cause given.
Own software
Nov 16
Optimizely hijack plants card-skimming code on 178 organisations' sites
Malicious code was inserted into 188 customer projects on Optimizely's website-testing service, so trusted scripts served from Optimizely's own network ran it on the sites of the BBC, CNN, Fox News, Zoom, Salesforce, GitLab, HSBC, Equifax, IKEA, Air France-KLM and more. It swapped crypto wallet addresses, skimmed card details typed into site forms, stole cookies and pushed exploit kits at older iPhones, for about 17 hours. How the attacker got in is not yet public.
Who felt it: anyone who typed card details into an affected site's own form that day. About $7,040 in crypto reached the attacker's wallets; card losses are unknown.
Confirmed attack
Nov 17
NSA Net, the agency's classified internal network, goes down
Six people familiar with the matter told CNN the network was down and staff arrivals were staggered. No evidence of a cyberattack; no official cause.
Cause not public
Nov 17
Apple: iCloud, Apple Pay, App Store falter
Apple says "an internal update" caused the outages and ruled out a cyberattack. Most services were back by the afternoon.
Own update
Nov 17
"Java4Shell": an AI-found critical flaw in Java
Start-up Calif disclosed CVE-2026-114938, a memory bug in Java's Base64 decoder on x86-64 servers that can let an attacker run code without logging in. It says an AI coding agent found it in about an hour for under $200. Oracle and OpenJDK shipped fixes the same day; CISA urged immediate patching. Not linked to any outage.
AI-found flaw

In the background: the FBI's 13 Nov warning about ARTEX (section 6), October's ConnectWise Cloud Backup compromise, and Barracuda's 44-hour cloud intrusion that began on 1 Nov with a forgotten vendor key.

BROKE THEMSELVES Microsoft · months-old hidden bug FAA · router during upgrade Chase · software bug Apple · internal update NSA Net · cause not public ATTACKED OR EXPOSED Fugaku · intruder, data taken Optimizely · skimmer, 178 orgs Java4Shell · AI-found, patched ARTEX · FBI warning No link found between the two groups
KKReckons graphic. Causes as stated by each company or agency, 12–17 Nov 2026.
2 · Anatomy of a cascade

How a careful company still lost a day of email

Microsoft's root-cause summary, shared by executive Charles Lamanna, is worth reading closely, because it undercuts the idea that this week was about carelessness. The faulty code was not rushed. It went through Microsoft's staged, safe rollout from July and ran worldwide for more than three months without trouble. Its labs simply never tested the right conditions: they focused on older, smaller servers, while the bug bit hardest on newer machines with more processor cores.

THREE QUIET CONDITIONS… July: a fix with a hidden bugFine for 3+ months; lab tests missed it −15 h: Teams moves onto the gatewayA normal failover; extra load, new traffic mix ~14% of capacity already offlineUnrelated fault the day before …THEN THE DOMINOES US peak traffic: servers lock upThe July bug finally triggers Safety net spreads the overloadTraffic rerouted to healthy servers, which lock up too <2 hours: 80%+ unresponsiveWalls between regions fail to hold Restart stormRevived servers swamped by waiting traffic Old code, no quick rollback: ~24 hours
KKReckons graphic based on Microsoft's root-cause summary. The final post-incident review is due later this week.

Two details stand out for anyone who runs systems. First, the safety feature did the damage: automatic rerouting, which normally saves the day, carried the overload from server to server. Second, the recovery tools assumed a few servers would fail at a time, not most of them, and because the bad code was months old, the usual one-click rollbacks couldn't undo it. Engineers had to patch and restart servers in big batches. The FAA's failure had the same shape: one change at one remote site, then bad data travelling the whole network.

The diagram shows why one company's bad day became everyone's. Many Microsoft services pass through the same gateway. Those that could switch to a backup came back sooner; email could not.

ONE DOOR, MANY ROOMSShared networkgateway80%+ servers frozeOutlook / ExchangeTeamsOneDriveSharePointOffice appsCouldn't move off: email down ~24 hoursMoved to a backup gateway: back soonerBehind the door: ~430 million users, 70+ countrieshospitals, courts, schools, banks, airlines
KKReckons graphic based on Microsoft's root-cause summary; user figures from published estimates.
3 · We've been here before

The giant outages almost always start at home

Big, shared systems fail often, and the worst failures of recent years were nearly all self-inflicted changes, not attacks. Cloudflare counted 174 major internet disruptions around the world in 2025 alone, more than three a week.

Jan 2023
FAA's pilot-warning system fails
A contractor deleted files from a key database; the first nationwide ground stop since 9/11 followed, and about 11,000 flights were disrupted.
Jul 2024
CrowdStrike's bad update crashes 8.5 million Windows PCs
Airports, hospitals and banks went down. Insurer Parametrix put direct losses for Fortune 500 firms at $5.4 billion.
Oct 2025
AWS and Azure, nine days apart
An Amazon database naming fault cascaded to 113 services for more than 15 hours; a Microsoft configuration change then disrupted Office 365, Xbox and Alaska Airlines for about nine.
Nov 2025
Cloudflare errors across the web
A company that carries about a fifth of internet traffic served error pages worldwide for hours.
Oct 2026
Google Cloud outage hits UK banks
Three months after the Bank of England began supervising "critical" cloud suppliers.
Nov 2026
This week
Microsoft, the FAA, Chase, Apple, NSA Net: four owned causes, one undisclosed.

The lesson from the record: a busy week of outages is normal for a world that runs on a few giant platforms. What has changed is how many people each failure reaches, and how quickly the attack side is moving.

4 · The rumours

How a one-minute alert test became the end of the world

In September, the FCC and FEMA announced a nationwide test of the Emergency Alert System for 2:20 pm Eastern on 17 November, the first since October 2023. It was meant to interrupt US television and radio for about a minute; phones would stay silent.

Over the following weeks, posts on TikTok, X, Instagram and Facebook turned the date into a "great blackout". Some promised 90 days without power, internet or banking; some tied it to a Pentagon project; France 24's fact-checkers traced the spread through cryptocurrency accounts, misread European advice on keeping 72-hour emergency kits, and the usual claims that The Simpsons had predicted it. Then last week's real outages arrived, and every one looked like confirmation.

The rumour also landed on prepared ground. Newsweek counts more than 30 governments advising households to keep emergency supplies, from 72 hours in Britain and France to up to ten days in Germany, and Europe has spent the autumn dealing with real sabotage (next section). When officials are telling you to stock a cupboard, a doomsday date feels less absurd.

HOW MANY DAYS TO PREPARE FORGermanyup to 10 daysSwedenabout 7 daysNorwayabout 7 daysUK3 daysFrance3 daysNetherlands3 daysCanada3 daysJapan3 daysThe US advises "several days"; 30+ governments give advice.
KKReckons graphic. Household emergency-supply advice by country, from Newsweek's survey of official guidance (October 2026).

CNN's media desk argued that feeds which reward alarm turned ordinary tech failures into signs of collapse. It also debunked smaller spin-offs, such as a Costco outage that didn't happen and a secret grid test that doesn't exist. Officials didn't always help: one cabinet member said on the day of the ground stop that nothing could be ruled out.

HOW A TEST BECAME A DOOMSDAYA routine test17 Nov, 2:20 pm ET, ~1 minute“90-day blackout” postsTikTok, X, Instagram, FacebookAmplifiers pile incrypto accounts, Simpsons memesReal outages hiteach one looks like proofFeeds reward alarmmillions see the doomsday
KKReckons graphic, based on fact-checks by France 24, PolitiFact and CNN. Each step reached more people than the last.

Now test yourself. Tap each claim to see how it holds up.

"It's a coordinated attack by a foreign state."
Not supported

Every affected company and agency says it found no sign of an attack on its outage. Director of National Intelligence Jay Clayton told CNBC the government had looked closely and ruled out both a cyberattack and a coordinated nation-state effort; National Cyber Director Sean Cairncross said the same. No group has claimed responsibility, and the causes given are different in kind at different firms.

"The world goes dark on 17 November."
Hoax

The only planned event was the alert test, announced by the FCC on 16 September with a backup date of 3 December. No government or grid operator announced a blackout.

"The FAA's new AI system grounded the planes."
Not supported

The FAA had expanded an $875 million AI tool for controllers, called SMART, days earlier. Transportation Secretary Sean Duffy said it played no part; the FAA's account points to the L3Harris router.

"Banks and the NSA went down too."
True

Chase confirmed a nearly eight-hour disruption on 16 Nov and blamed a software bug. CNN reported NSA Net down on 17 Nov, with no evidence of an attack. Real, but not shown to be connected.

"AI is finding three times more flaws than last year."
Roughly right

Public vulnerability records (CVEs) passed 91,000 with two months of 2026 left, nearly double all of 2025, and critical ones are about 2.4 times last year's total, according to Akamai. AI is a main driver: Anthropic says partners in its Project Glasswing found at least 129,000 verified vulnerabilities between April and July.

"Exploits now arrive before fixes: minus seven days."
Real, often misread

Mandiant's M-Trends 2026 put the mean time to exploit for 2025 at −7 days: among flaws it saw exploited, attacks often began before a patch existed. It's an average over exploited flaws, not all bugs. For flaws on CISA's exploited list, the median gap from publication to confirmed attack has fallen from 33 days in 2022 to about 6 in 2026.

"…so rushed security patches caused this week's outages."
The weak link

No firm has said its failing change was a security patch. Microsoft's faulty fix was the opposite of rushed: it rolled out carefully over weeks and ran for months. The FAA's was a years-long copper-to-fibre swap. The doom loop describes real pressure in the industry, not a proven cause of last week.

THE VIRAL LOOP, FACT-CHECKED 1 · AI finds more flaws~2× CVEs vs 2025 (Akamai) 2 · Exploits beat fixes−7 days mean time to exploit (Mandiant) 3 · Pressure to patch faster3-day US federal rule (CISA BOD 26-04) 4 · Rushed changes break thingsNot shown for this week's outages 5 · Fixes lag5 months 10 days for criticals (Qualys) Dashed: the step the viral post asserts butthe evidence does not yet show
KKReckons graphic. Sources: Akamai; Mandiant M-Trends 2026 via ISACA and CSA; CISA BOD 26-04 via CSA; Qualys 2026 benchmark via CSA.
5 · The real attacks, in Europe

The cold war in the freezer aisle

If you want evidence of hostile states hitting ordinary infrastructure, it isn't in last week's American outages. It's in Europe's supermarkets.

Over the night of 11–12 October, refrigeration controllers in hundreds of French supermarkets (up to 700 by some counts, across Intermarché, Netto, Coopérative U, Carrefour, Monoprix and others) were switched into a defrost cycle with the food still inside. With no alarms and few staff on site, many cases reached about 50°C within six hours; one Normandy store reported 80°C. Some units melted or caught fire, and stores threw out whole frozen aisles. On 2 November it happened again, this time across Belgium, Luxembourg, the Netherlands, Sweden, Estonia, Ireland and beyond: Delhaize, Carrefour, Colruyt, Coop, SuperValu, an IKEA store and others reported more than 100 sites hit in a day. Hackers also tried to heat refrigerators at three French blood banks; monitoring alarms let staff step in before plasma was lost.

WHERE THE FRIDGES WERE HITSwedenEstoniaIrelandUKNeth.GermanyBelgiumLux.FrancePortugalFranceup to ~700 stores, 11–12 Oct;3 blood banks targeted 2 NovFirst wave, 11–12 Oct (blamed on Iran by France)Second wave, 2 Nov: 100+ sites in a dayA small number of UK cases (NCSC)
KKReckons tile map: each square is a country, placed roughly where it sits in Europe, not to scale. Some reports also mention Italy, Poland and Spain; we could not confirm them.
HOW A FRIDGE GETS HACKED Controller left on the open internetIts web page reachable by anyone Username "admin"; password = today's dateA hidden "code of the day" bypass Old, unpatched kitSome models unsupported since 2016 and 2021 Overnight: switch to defrostNo alarm fires; few staff on site ~50°C in six hoursSpoiled stock, melted units, some fires;tens of millions of euros in losses
KKReckons graphic. Flaws as reported by researchers at Claroty Team82 and security researcher Kevin Beaumont. The flaws were disclosed in March 2025; Danfoss issued a fix for newer controllers that August, while older models no longer get updates.

Who did it? On 9 November France's foreign minister, Jean-Noël Barrot, blamed the October attacks on Shahid Kaveh, a cyber unit tied to Iran's Revolutionary Guards. He called it the first large-scale cyber sabotage on French soil. A joint report by French cyber and intelligence agencies cited Iran-linked infrastructure and IP addresses, traces of Farsi and the unit's known methods. Paris summoned Iran's ambassador, sought EU sanctions, and prosecutors requested international arrest warrants for the unit's leader and two deputies; Barrot said France could respond with cyber measures of its own. Reuters called the hacks among the most disruptive since the Iran war began in February.

The 2 November wave has not been formally attributed. Italy's defence minister, Guido Crosetto, said groups in Iran, China and Russia had been identified, while stressing that cyberattacks are hard to attribute and governments may not be directly involved. Some researchers urge caution the other way: because the flaw is so easy to use, Ken Munro of Pen Test Partners suggested copycats or "kids messing around" could be behind some incidents. Across the Atlantic, the Pentagon is treating freezer failures at US military commissaries as suspected cyberattacks, and on 5 November CISA, the FBI and three other US agencies warned of ongoing threats to the food cold chain.

Why this matters more than last week's outages

This was deliberate sabotage of everyday life, using no AI at all: just old devices, exposed to the internet, with passwords anyone could guess. Britain's cyber chief, NCSC head Richard Horne, said his team worked through the night to warn UK firms, and called the fridge attacks "just the foothills". His agency expects AI-powered attacks on British critical infrastructure by 2028, because AI will let attackers find and exploit long-tolerated weaknesses at enormous scale. Firms that "got away with tech on the cheap", he warned, will have to explain empty shelves and dark lights.

Britain's defence secretary, Wes Streeting, went further on 22 October. He asked the public to prepare for what he prefers to call "state-sponsored terrorism" against everyday infrastructure, such as cyberattacks that knock out business payment systems or collapse supply chains, and the government issued a three-day "Be Ready" checklist. Days later he wrote that Britain has little more than 1,000 days to prepare for possible conflict with hostile powers by 2030.

6 · The real worries

Not one big hack. A change in speed.

Strip out the rumours and five real trends remain. None caused last week's outages, but each makes the next bad week more likely.

1. AI now runs whole break-ins

On 13 Nov the FBI warned that criminals and state-aligned hackers are using ARTEX, an open-source tool released on GitHub in July by a developer called "Autumn-27". You describe a goal in plain language and a team of AI agents does the work, from scanning to stealing data, at what the FBI calls "machine speed". Some intrusions reached admin-level access in under three hours. Google's threat team, with Mandiant, watched it split work across six specialised agents and log 27 actions in 23 seconds, a pace no human hacker keeps.

INSIDE AN AI-RUN ATTACK Operator types a goal"Get into this company's cloud" Planner agentBreaks the goal into tasks Worker agentsRun real hacking and admin tools Shared map of the targetEach finding wakes the planner LOOPS IN SECONDS Admin access in <3 h, then data outLow-skill actors can run many at once
KKReckons graphic, simplified from descriptions by the FBI, Google Threat Intelligence and independent analysts.

ARTEX doesn't contain an AI model; it plugs into commercial ones. Attacker-run copies mostly use cheap Chinese models such as DeepSeek, though the FBI lists Western providers among those it can connect to. The developer pulled it from GitHub in October after abuse reports, but forks and a paid "Pro" version keep it alive, and researchers have found more than 1,600 exposed control panels. Microsoft says a China-aligned group it calls Storm-0940 used it in October's ConnectWise Cloud Backup compromise, attempting sign-ins at more than 2,250 downstream Microsoft 365 customers (attempts, not confirmed breaches). In South Korea, ARTEX attacks on financial firms took personal data on about 67,000 people.

2. Finding flaws now far outruns fixing them

Java4Shell shows both sides. A researcher pointed an AI coding agent at the most delicate part of Java's engine and had a critical, remotely exploitable bug in about an hour, then a working exploit in roughly two more. Java runs at most big banks, airlines and government agencies. Defenders found this one first. But the fixing side is buckling: about 300 new CVEs a day, 2,308 in the Linux kernel alone in the first half of the year, and, by Anthropic's count, open-source maintainers had patched only 516 of the 6,157 vulnerabilities it had reported to them by early October. Google's Project Zero now says AI makes old disclosure timetables "clearly insufficient".

THE SQUEEZE IN THREE NUMBERSMore flaws: new CVEs per day2020~502026~300Faster attacks: days to first exploit*202233 days2026~6 daysSlow fixes: open-source flaws from AnthropicReported by early Oct6,157Patched516
KKReckons graphic. *Median, publication to confirmed exploitation, for flaws on CISA's exploited list. Sources: Akamai (vendor) for CVE volume and exploit timing; Anthropic via Software Improvement Group for reported and patched flaws.

3. Trusted suppliers are the new front door

Optimizely, ConnectWise and Barracuda share a pattern: the attacker came in through a supplier the victim already trusted. Barracuda, itself a security company, says its intruder used a cloud key it had given an AI vendor and never retired; the vendor's own systems were compromised in October. Scott Helme, whose team spotted the Optimizely attack, put the lesson in a line: trusted code "doesn't always remain trustworthy".

THE TRUSTED BACK DOORAttackerplants code in Optimizely customer projectsOptimizely's own networkserves it as normal, trusted script178 organisations' websitesBBC, CNN, Zoom, HSBC, IKEA, Equifax…Your browser runs itno warning; it looks like the real siteSkims carddetailsSwaps cryptoaddressesAttacks oldiPhonesLive for about 17 hours on 16 November
KKReckons graphic based on Report URI's analysis. How the attacker got into Optimizely is not yet public.

4. Secrets leak and get used in seconds

Stony Brook University researchers published 300 decoy container images holding planted cloud and AI keys. On Docker Hub, the fastest AWS key was used 7.86 seconds after publication, the median within 51 seconds, and all 100 AWS keys within 15 minutes. Planted OpenAI and Anthropic keys were drained too. Most of the bots never ran the containers; they just unpacked them looking for secrets.

5. Change itself is the risk, and the stakes are concentrated

Four of last week's five outages trace back to their owners' own software or hardware work. FAA chief Bryan Bedford said every government agency has flagged the FAA's telecom network as a "critical weakness" for more than two decades; it costs about $1 billion a year just to keep running. Former Bank of America operations chief Sumeet Chabria called big upgrades "fixing the plane while it's flying", and said last week that stopped being a metaphor. Because so much now runs on a handful of providers, one bad change reaches hundreds of millions of people.

HOW FAST IT MOVES · log scale 1 min 1 hr 1 day 1 mo 1 yr Leaked AWS key used8 sec AI finds Java4Shell~1 hour ARTEX: target → admin<3 hours Barracuda intrusion44 hours Published → exploited (median)~6 days Fix a critical flaw (mean)~160 days Red: attacker speed. Teal: defender speed.Each gridline is a big jump, so the gap is wider than it looks.
KKReckons graphic. Sources: Stony Brook ContainerBots study via BleepingComputer; Calif via Unite.AI and Tech Startups; FBI FLASH-20261113-001; Barracuda; Akamai (2026 median); Qualys 2026 benchmark via CSA (5 months 10 days ≈ 160 days).
7 · The debate

Two camps, and where they agree

"It's operations, not Armageddon"

Officials, the companies and many engineers. The causes are mundane and different; big outages have always come in clusters; the record of the past three years is overwhelmingly self-inflicted. Defenders have the strongest AI tools today, and early research on Project Glasswing finds that access helped firms surface more flaws without lowering quality.

"The clock has changed"

Threat researchers, the FBI and many security chiefs. Attack tools that once needed experts now need a sentence. Flaws are found faster than they can be fixed; supply-chain doors are everywhere; keys are stolen in seconds. A week of honest mistakes shows how fragile things are before an attacker deliberately pushes, and Europe's fridges show attackers already are.

Where they agree: the outages were not a coordinated attack; concentration on a few providers turns small mistakes into global events; and the answer is engineering discipline and faster fixing, not panic.

Reasons for calm

  • Clusters happen: 174 major internet disruptions in 2025.
  • Mundane causes: a router, a hidden bug, another bug, an update.
  • Defenders have AI too: Java4Shell was found and patched by the good side the day it went public.
  • Basics work: sites with strict script rules blocked the Optimizely code.

Still unresolved

  • Microsoft's final review is due this week.
  • The FAA hasn't said if it was human error, installation or hardware.
  • The NSA has given no cause; Apple and Chase few details.
  • How Optimizely was breached, and who hit Fugaku, are unknown.
  • Early "not an attack" statements can change once logs are read.

Who says what

ReassuringMixedWarning
Jay Clayton
Director of National Intelligence

The government looked closely; it was not a cyberattack and not "a coordinated effort by a nation state".

Sean Cairncross
White House National Cyber Director

The ground stop and Monday's outages were not a coordinated, malicious attack.

Bryan Bedford
FAA Administrator

The failing network has been flagged as a "critical weakness" across government for 20+ years.

Sean Duffy
US Transportation Secretary

Without a modern design, failures like this will keep happening; he wants $30 billion more.

Sumeet Chabria
CEO, ThoughtLinks; ex-Bank of America operations chief

Big upgrades are "fixing the plane while it's flying"; the lesson is concentration risk.

Chris Nyhuis
CEO, Vigilant

An attack was possible, but people shouldn't jump to conclusions in either direction.

Richard Horne
CEO, UK National Cyber Security Centre

The fridge attacks are "just the foothills"; expect AI-powered attacks on UK infrastructure by 2028.

Jean-Noël Barrot
French Foreign Minister

Called the fridge hacks France's first large-scale cyber sabotage and blamed Iran's Revolutionary Guards.

Wes Streeting
UK Defence Secretary

Britain should prepare for "state-sponsored terrorism" against everyday infrastructure.

Guido Crosetto
Italian Defence Minister

Groups in Iran, China and Russia were identified, but attribution is hard and states may not be directly involved.

Ken Munro
Pen Test Partners

The fridge flaw is so easy that copycats or "kids messing around" could be behind some attacks.

FBI
FLASH alert, 13 Nov 2026

ARTEX automates intrusions "at machine speed" and lets low-skilled actors run complex attacks.

Quang Luong & Bruce Dang
Researchers, Calif

Java sits at the core of the economy; an AI agent found a critical flaw in it in about an hour.

Scott Helme
Founder, Report URI

Trusted third-party code "doesn't always remain trustworthy".

Google Project Zero
Vulnerability research team

AI-speed exploitation makes old disclosure timetables insufficient.

Billy Tsouvalas & Nick Nikiforakis
Stony Brook University

Their decoy keys were found and used within seconds, mostly by bots.

Linux stable-kernel maintainer
via Akamai

Expects the AI-driven flood of bug reports to last at least another 18 months.

Quotes are short and attributed; all other wording is KKReckons' summary of the source.

8 · How it can be controlled

What actually reduces the risk

No single defence is perfect; each has holes. Safety engineers stack several, so a mistake or attack that slips through one is caught by the next. Microsoft's outage got through because several holes lined up at once: a hidden bug, extra load, missing capacity and region walls that failed.

LAYERS, NOT A WALLSmall, staged changesWalls between regionsKey and vendor hygieneOff the open internetMonitoring + AI defenceRed: holes line up, the failure or attack gets through.Teal: one solid layer stops it.
KKReckons graphic, after the "Swiss cheese" model of safety used in aviation and medicine.

For companies and engineers

  • Make changes small and reversible. Roll out to a few servers or sites, watch, then widen; keep a fast rollback. The FAA's problem stopped once the new site was blocked.
  • Test where it hurts, and plan for mass failure. Microsoft's lessons: load-test on the newest, biggest hardware at peak traffic; make sure walls between regions really stop a cascade; and build recovery that can restart a whole fleet, not just a few servers.
  • Kill single points of failure. Map which critical work depends on one provider, and keep a tested plan B for email, payments and communications.
  • Patch by exposure, contain first. With attacks measured in days, fix what attackers can actually reach and fence off exposed systems while patches are tested. Monthly patch cycles are now slower than the median attacker.
  • Treat keys like cash. Scan code, pipelines and container images for secrets; revoke, don't just delete, anything that leaked; retire vendor keys when the work ends.
  • Follow the FBI's ARTEX advice: list every internet-facing app, keep admin panels off the open internet, limit what third-party apps can do, watch for new credentials and permission grants, rate-limit sensitive APIs, and stop showing detailed errors to strangers.
  • Lock down third-party scripts with a strict Content Security Policy that avoids unsafe-eval, and monitor for new scripts and data destinations.
  • Get machines off the open internet. Fridges, building controls and factory kit belong behind private networks, with default passwords changed, alarms on and unsupported devices replaced. The fridge attacks used flaws public since March 2025, on devices that were never updated or no longer supported.
  • Use AI for defence. Run the same agents attackers use against your own code first, and on a schedule.

For governments and regulators

  • Fund the boring upgrades. Congress approved $12.5 billion for air traffic control last year; Duffy wants $30 billion more, and the FAA plans to move controller communications from analogue to digital in late 2027.
  • Set patch clocks to match the threat. CISA's June directive, BOD 26-04, gives federal agencies three days to fix the highest-risk flaws.
  • Supervise critical suppliers. The Bank of England began overseeing its first designated critical cloud providers in July; watching a few platforms targets concentration risk directly.
  • Ban weak defaults and name the attackers. Britain has banned default passwords on new consumer connected devices since 2024, and similar rules could extend to business kit; France's public attribution, sanctions push and arrest warrants show the price states are now willing to impose.
  • Help the fixers. Open-source maintainers can't keep up with AI-found bugs; funding, staff and AI tools for patching matter as much as tools for finding.
  • Explain outages fast and fully. Quick, detailed post-mortems starve rumours. Microsoft's root-cause note within days is the model.

Your 10-minute resilience check

  • Update your phone and laptop. The Optimizely code targeted older iPhones; Report URI advises iOS 18.7.3 or 26.3 or later.
  • Check card statements if you typed card details into a big brand's own site form on 16 Nov.
  • Double-check any crypto address before sending; the attack swapped them on screen.
  • Keep a second card or bank account, a little cash, and the 72-hour basics many governments now advise: water, non-perishable food, medicines, a torch, batteries and a radio.
  • Save boarding passes, tickets and key documents offline.
  • Turn on passkeys or app-based two-step sign-in for email, bank and cloud accounts.
  • Before sharing a scary post, check the company's or agency's own status page.
9 · What to watch next

Six things that will tell us more

  1. Microsoft's final post-incident review, due this week: did anything else go wrong, and what changes?
  2. The FAA and L3Harris findings: human error, bad installation or faulty hardware, and whether the copper-to-fibre work pauses.
  3. How Optimizely was breached and how many card numbers were taken.
  4. Java4Shell in the wild: whether attackers start exploiting unpatched servers before companies update.
  5. Europe's response to Iran: whether the EU adopts sanctions over the fridge attacks, and who is blamed for the 2 November wave.
  6. ARTEX copies: new forks, new victims, and whether AI providers can shut off attackers' access.
Bottom line

The world isn't ending. It's getting faster.

Last week's outages were ordinary failures of change, in systems too big and too shared to fail quietly. The doomsday was a rumour built on an alert test. The genuine state sabotage was in Europe's supermarkets, and it needed nothing cleverer than a default password. The new thing ran alongside: AI that finds and exploits weaknesses in hours, keys stolen in seconds, and attackers who walk in through suppliers we trust. Neither story needs a conspiracy. Both need the same response: safer changes, fewer single points of failure, tighter control of access, and defenders who use AI first.

The takeaway

The danger isn't one giant hack. It's many small breaks in systems too complex to fix fast, while attackers speed up.

Correction to our 17 November visual explainer. We marked the US banking disruptions and the NSA network outage as "in post, not verified". Both have since been reported: Chase confirmed a software-bug outage on 16 Nov, and CNN reported NSA Net down on 17 Nov. The explainer's "~12,000 flights cancelled" was, if anything, low: FlightAware counted more than 9,000 cancellations on Friday and over 4,000 more by Saturday morning. The explainer's conclusion stands.
Sources
KKReckons.com · by KK Reddy
Figures come from primary documents and independent reporting, in our own words. Vendor and think-tank figures are labelled. If we could not confirm it, it is not here. No sponsored content. All graphics are KKReckons originals.