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.
Outlook stops loading. Within two hours, more than four in five of the servers in front of Microsoft's email have frozen.
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.
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.
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.
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.
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.
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.
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.
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.
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.
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.
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.
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.
Now test yourself. Tap each claim to see how it holds up.
"It's a coordinated attack by a foreign state."
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."
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."
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."
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."
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."
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."
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 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.
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.
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.
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".
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".
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.
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.
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
The government looked closely; it was not a cyberattack and not "a coordinated effort by a nation state".
The ground stop and Monday's outages were not a coordinated, malicious attack.
The failing network has been flagged as a "critical weakness" across government for 20+ years.
Without a modern design, failures like this will keep happening; he wants $30 billion more.
Big upgrades are "fixing the plane while it's flying"; the lesson is concentration risk.
An attack was possible, but people shouldn't jump to conclusions in either direction.
The fridge attacks are "just the foothills"; expect AI-powered attacks on UK infrastructure by 2028.
Called the fridge hacks France's first large-scale cyber sabotage and blamed Iran's Revolutionary Guards.
Britain should prepare for "state-sponsored terrorism" against everyday infrastructure.
Groups in Iran, China and Russia were identified, but attribution is hard and states may not be directly involved.
The fridge flaw is so easy that copycats or "kids messing around" could be behind some attacks.
ARTEX automates intrusions "at machine speed" and lets low-skilled actors run complex attacks.
Java sits at the core of the economy; an AI agent found a critical flaw in it in about an hour.
Trusted third-party code "doesn't always remain trustworthy".
AI-speed exploitation makes old disclosure timetables insufficient.
Their decoy keys were found and used within seconds, mostly by bots.
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.
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.
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.
Six things that will tell us more
- Microsoft's final post-incident review, due this week: did anything else go wrong, and what changes?
- The FAA and L3Harris findings: human error, bad installation or faulty hardware, and whether the copper-to-fibre work pauses.
- How Optimizely was breached and how many card numbers were taken.
- Java4Shell in the wild: whether attackers start exploiting unpatched servers before companies update.
- Europe's response to Iran: whether the EU adopts sanctions over the fridge attacks, and who is blamed for the 2 November wave.
- ARTEX copies: new forks, new victims, and whether AI providers can shut off attackers' access.
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 danger isn't one giant hack. It's many small breaks in systems too complex to fix fast, while attackers speed up.
- FBI FLASH-20261113-001, ARTEX: ic3.gov
- ARTEX analysis: Google Threat Intelligence; Zscaler; SOCRadar; ExtraHop
- Microsoft 365 outage: root-cause summary posted by Charles Lamanna on X; IncidentHub; Wikipedia; CNN ($7bn estimate); Forbes (NHS)
- FAA ground stop: FAA Q&A; NPR; Washington Post; TechCrunch; AeroCorner; BankInfoSecurity; Reuters via Fox 13
- Fugaku: RIKEN via HPCwire; WindowsForum; Japan Cyber Watch
- Chase and NSA Net: Tom's Guide; Gulf News; CNN via WRAL
- Apple: UPI; 9to5Mac
- Officials on attacks: CNBC; Yahoo Tech; WMTV
- Blackout hoax: FCC notice DA 26-989; France 24; PolitiFact; CNN
- Optimizely: Report URI; BleepingComputer; Malwarebytes
- Java4Shell: New York Times (paywalled); free: Unite.AI, Lynnwood Times, Tech Startups
- ConnectWise, Barracuda: Dark Reading; Barracuda
- Docker Hub keys: BleepingComputer; We Fix PC
- Vulnerability data and AI defence: Akamai (vendor); ISACA; CSA; CSA patch wave; CSA on Project Zero; CSA on BOD 26-04; SIG on Glasswing; Queen Mary study (arXiv)
- Past outages: FAA 2023 via Reuters; CrowdStrike/Parametrix; AWS, Azure and Cloudflare Radar data; Cloudflare 2025; Google Cloud and UK banks
- European fridge attacks: Guardian; CNBC; Al Jazeera; AP via ABC; Wikipedia; France blames Iran; Reuters via Al-Monitor; Axios
- NCSC warning: Reuters via Insurance Journal; AP; City AM
- Preparedness: Guardian on Streeting; Infrastructure Now; UPI; Newsweek stockpiling map
- Concentration risk: Yahoo Finance / American Banker; Science Media Centre
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.


