In brief…
An investigation of 193 significant digital incidents between January and 14 November 2026 found that malice was the largest single identified cause, but accounted for fewer than half of the incidents analysed. Of 48 significant data compromises, 44 (91.7%) were malicious. But for 58 of 140 operational disruptions (41.4%), the available public evidence did not establish a cause. Case studies suggest that the consequences of disruption depend not only on what starts an incident, but also on whether safeguards, fallback systems and recovery arrangements work.
=====
Friday 13th. The day was almost over. I had one last job to do, to check the news. My eyes were almost shut. Then I saw it.
Planes grounded across the United States. The Federal Aviation Administration (FAA) had ordered a nationwide ground stop after a communications failure affected its air traffic control centres.
Unfortunately, it wasn’t the only failure that day. Since Thursday evening, a Microsoft outage had left businesses, hospitals and government agencies around the world struggling without Outlook and Teams. Neither had been publicly attributed to an attacker. But, as I’ve since discovered, the first explanation isn’t always the final one.
I stared at the words I’d written a few hours before in my last post on LinkedIn …
“Malice. Mistake. Malfunction.”
Having just gone into my 30th year in cybersecurity, I know how quickly our attention turns to malicious actors when something goes wrong. Everyone wants to know who attacked us? How did they get in? What did they take?
But what happens when nobody attacks?
It’s a question I’ve been asking more and more as I explore what makes organisations capable of surviving a digital crisis in the AI era.
It got me thinking. Of this year’s significant digital incidents, how many could be attributed to each? Could that be established from the public record alone? Wouldn’t it be interesting to find out?
My journalistic instincts kicked in. Goodbye Saturday and all my chores. Hello research!
What I set out to find
I didn’t set out to prove a point. I genuinely didn’t know what I’d find, and that was the appeal. It was an adventure in search of the truth, or at least as close to it as the public record, the tools available to me, and the time I had would allow.
I used AI research assistants to help investigate more than a thousand publicly reported incidents, working to a fixed set of rules. They searched for consequences rather than causes, checked reports against available sources and proposed classifications, which were then challenged where the evidence was uncertain. It wasn’t a conventional academic study, but I wanted to see how close I could get to a credible answer using public information, and AI tools. In the end, 193 incidents between January and mid-November met a fixed test of significance, including this week’s FAA ground stop and Microsoft outage, which I added once the main search was complete.
The first finding
Malice was the largest single identified cause – 87 of the 193 incidents, or 45%. Mistakes accounted for 17, malfunctions for 19 and external events, such as a failed utility cable, for 10. For the remaining 60, the public sources I examined did not establish a cause.
Malice led, but it was not a majority, and that figure hid the most important finding of all.
The second finding
When I separated the two kinds of incident, they looked almost nothing alike.
- Data breaches are a story of malice. 44 of 48 significant data compromises, more than nine in ten, were attacks – theft, extortion, intrusion.
- Disruptions are a story we’ve not been told. Of 140 operational disruptions, 38 were attacks and 44 were mistakes, malfunctions or external events. For the remaining 58, more than four in ten, the public sources I examined did not establish what caused them.
Unknown does not mean non-malicious. Some of these disruptions may have been attacks; others may have been mistakes, technical failures or external events. The public evidence doesn’t establish which.
In other words, Malice leads the overall count largely because of the data breaches. For the outages that stop hospitals, banks and networks, the public record cannot tell us which cause leads.
The cause we cannot see
However, we can see that the 58 unexplained disruptions are concentrated in particular sectors. 22 are payment and banking outages, and 13 are telecom outages – sectors where operators typically restore service, apologise and move on. In government and health, by contrast, most disruptions were attacks, and attacks get announced.
An attack is newsworthy and may trigger mandatory disclosure. A botched update or failed power supply is embarrassing, but often attracts far less scrutiny and may never require public explanation, even when the disruption is significant. Public reporting may therefore over-represent malice among the incidents whose cause we know. If every one of the 58 were an attack, attacks would make up 69% of disruptions; if none were, 27%. The public evidence cannot narrow that range.
Regulatory reports published in 2026 offer a useful comparison, although they cover incidents from 2025. EU financial firms reported 3,383 major ICT incidents that year, with system failures accounting for around half and cybersecurity incidents just 10%. The European Central Bank also found that 38% of banks’ major incidents had an IT change as their root cause.
These are different datasets, collected under different reporting rules, so they cannot tell us what caused the 58 unexplained disruptions in my investigation. But they do suggest that mistakes and technical failures may be more prevalent than public reporting alone would lead us to believe.
Even mandatory reporting has gaps – a quarter of the resilience incidents UK telecom providers reported to Ofcom had no cause recorded.
Why mistakes stay quiet
There may also be a cultural dimension to this silence. When an organisation is attacked, it can describe itself as a victim. When a disruption follows a failed change or an expired certificate, the questions turn inward – who approved it, who tested it, and why didn’t the safeguards work? That can make disclosure uncomfortable. But the public record rarely tells us whether silence reflects embarrassment, uncertainty, legal constraints or something else.
Whatever the reason, each time a cause goes unpublished, everyone else loses a lesson they could have learned cheaply, and our industry’s favourite story grows stronger – that the danger is out there, rather than in how we build, change and run our own systems.
What happens after the first failure?
In October, part of Google Cloud’s London region went dark. A utility cable failed. That alone should not have mattered, because the site had a second power feed. But the second feed was out for repair. Then the generators that should have taken over didn’t start, because of a misconfiguration in their controls.
The first failure was not the story though. The story was what happened when each protection designed to contain it was needed.
Portugal’s national health service shows the same pattern. Its core systems ran from one data centre; a second, real-time site had been promised since 2022 and was not yet live. A power failure took e-prescribing down nationwide.
Shared suppliers turned single failures into many – 37 of the incidents name a third party, vendor or hosting provider in the cause or the weaknesses. Regulators see the same thing, with about a third of major incidents at EU financial firms originating at third parties.
Cause may also affect how long an outage lasts. Only eight of the 38 malicious disruptions have a sourced duration, but seven of them lasted four days or more; non-malicious outages with a sourced duration typically lasted hours. Anyone who has recovered from ransomware will recognise why – you cannot restore until you trust your immutable, tested backups. With so few cases, treat that as an indication, not a conclusion. And because public sources describe a failed safeguard for only 44 of the 193 incidents, the cases above are illustrations, not a measured rate.
None of this suggests malicious attacks deserve less attention. A destructive cyberattack can be far harder to recover from than a routine technical failure. The question isn’t whether organisations should defend against attackers. It’s whether they’re equally prepared for serious disruption when no attacker is involved.
Five questions your board should be asking
- For our last three significant outages, do we know, and have we recorded, what actually started them?
- When a significant outage occurs, do we investigate and share the lessons from internal mistakes and supplier failures as rigorously as we do from cyberattacks?
- Which of our critical services has a genuine second path, and when did we last prove that it works?
- Which single supplier or provider could take several of our services down at once?
- After a destructive attack, how long would it take us to trust our systems enough to restore them?
Where the conversation needs to change
Friday 13th reminded us that critical services don’t need an attacker to bring them to a halt. A communications failure can ground a nation’s aircraft. A problem deep inside one technology provider can leave governments, hospitals and businesses unable to talk to each other.
In two of the cloud incidents I examined, the provider’s early statement was later superseded by its own post-incident report. We should wait for the full accounts of this week’s failures, and then expect them to be published. Some cloud providers already publish detailed post-incident reviews. The banks, telecoms operators and public services that run critical systems should do the same, for their mistakes as well as for the attacks they suffer.
I set out to discover what really causes significant digital incidents. What I found, above all, was how little we can establish about why so many critical services fail.
We spend a great deal of time asking who caused an incident. We should; attribution can matter enormously. But we also need to ask whether the organisation can keep operating, whether its safeguards will actually work, whether its suppliers can recover, and who has the authority to act when normal systems fail.
The reason why is because when the next major disruption arrives, the initiating cause will matter, but so will everything that determines whether the organisation survives its consequences.
Now I want to hear from you…
Are we as willing to investigate and share lessons from mistakes and malfunctions as we are from cyberattacks? Join me on LinkedIn and tell me in the comments. I’ll be waiting for you there.
=======
PS. If you’re curious as to how I did this, here’s the detail.
- I classified incidents by their initiating cause: Malice: A deliberate act, such as a cyberattack, intrusion or sabotage. Mistake: A human or organisational action, such as a faulty configuration change, deployment or maintenance step. Malfunction: An unintended technical failure, such as a hardware fault, software defect or automation failure, without an identified triggering human change. I also recorded external causes, such as utility failures, separately. Where the available evidence did not establish a cause, I classified the incident as Unknown.
- Searching by consequence, not cause. 872 searches, month by month, across sectors and regions, looked for outages, breaches and service failures rather than “cyberattacks”. The approach was designed to reduce the tendency to find attacks more readily than other failures.
- Checking everything found. The searches produced 1,151 organisations to check, and every one was checked against available sources through the AI-assisted process, including vague-headline candidates that turned out to include dozens of hospital, council and payment outages.
- A threshold for “significant”. An incident counted if it met at least one fixed test: data on 100,000 people (10,000 for health data); a service down for an hour for 100,000 users or ten organisations; any hour-long interruption of a hospital, utility, bank, telecom, transport or government service; a safety effect; a US$10 million loss; or a government-confirmed state attack.
- Counting the start, not the headline. Incidents were dated by when they began; 337 reported in 2026 had started earlier and were left out.
- What was counted. 195 incidents met the test, including this week’s FAA ground stop and Microsoft outage, which were added after the main search using the same significance tests and coding rules (the Microsoft outage disrupted service for well over 100,000 users and government agencies; the FAA ground stop interrupted a national transport service for more than an hour); two precautionary website takedowns were set aside, leaving 193. Another 471 confirmed 2026 events had incomplete evidence of significance or timing, and 299 could not be verified. All are kept on record.
- Limits. These figures describe the incidents collected, not all incidents worldwide. Public information cannot reveal a cause an operator has not disclosed, and the sample leans towards North America and Europe. Classifications were proposed by AI agents and checked through an AI-assisted review process; they have not been independently validated by a human. As a final consistency check, a separate AI model coded a random sample of 30 incidents from a blinded evidence pack and agreed on 23 (76.7%, Cohen’s kappa 0.69), including all 10 attacks. Most disagreements involved incidents the register coded as Unknown, which shows how hard it is to tell an undisclosed fault from a genuinely unknown cause. The two most recent incidents are coded from early official statements and may change when full reports are published.

My full methodology, and supporting research is based on my independent, AI-assisted investigation of 193 significant digital incidents that began between 1 January and 14 November 2026. The analysis draws on publicly available evidence and uses defined criteria to classify initiating causes as Malice, Mistake, Malfunction, External or Unknown.
The findings reflect the incidents identified and verified through this investigation, not every digital incident worldwide. The supporting research notes explain the methodology, classification rules, limitations and sources.
Download the research notes and supporting evidence (PDF).
