I have moved cities. I have moved houses. My children were small and then, in what felt like the space of a long week, they were not. Somewhere in there the internet went from a thing you explained to your parents to a thing your parents use to book their own flights, and then AI arrived and we all started explaining things to our parents again.

Everything moved. Everything except the part I actually work on.

I do not mean nothing happened in security. Plenty happened. The budgets got bigger. The expense accounts got smaller. The conference floor got bigger than the budgets. And the technology is genuinely better than it was, in ways that would be hard to overstate to anyone who was not there for the earlier version of it.

What I mean is narrower, and it has been bothering me for a while. We got better at almost every technical thing we set out to get better at. And the threat got worse anyway. Not more sophisticated. Worse.

The part that actually changed

AI changed things. But the shape of it is familiar. Enormous capability arriving faster than governance, deployed by business units that did not ask permission, defended by people who inherited the problem rather than chose it. I have seen this movie. Most of us in this industry have seen it twice. It is innovation, it is running fast on both sides, and it is not the reason the loss numbers look the way they do.

The part that did not

Americans reported losing 12.5 billion dollars to internet crime in 2023. In 2024 it was 16.6 billion. In 2025 the FBI's Internet Crime Complaint Center recorded 20.877 billion dollars, a 26 percent increase in a single year, across 1,008,597 complaints, up from 859,532 the year before.

Losses among victims over 60 went from 4.885 billion in 2024 to 7.748 billion in 2025. The FBI's press release leads with 37 percent, which is the growth in complaints. The report itself puts 59 percent against the losses.

Detection is better than it has ever been. Response is better. Sharing is better. The losses went up anyway, and they went up steepest among the people least able to absorb them.

Not smarter. Crueler.

Ransomware payments actually fell in 2025, to around 820 million dollars, down 8 percent, according to Chainalysis. Good news, and most of the coverage treated it that way.

Read the rest of the page and it stops being good news. Claimed attacks rose about 50 percent over the same period. And the share of victims who pay anything has collapsed. Chainalysis puts it at 28 percent, a possible all time low. Coveware, which negotiates these cases for a living, recorded 23 percent in the first quarter of this year and a new record low in the second. Verizon's breach data puts it at 31 percent, which is another way of saying 69 percent of victims said no.

More attacks. Far fewer players. Now think about what that does to the business model.

Encryption stopped working as leverage, because backups and recovery got better, and our industry should take the credit for that. So the leverage moved. It moved to the one thing people will still pay to prevent, which is harm to a person.

You can watch it move. Google's Threat Intelligence Group found data theft in 77 percent of the ransomware intrusions it analysed in 2025, up from 57 percent the year before. Europol put it plainly in this year's threat assessment: the extortion has shifted from pressuring victims to get their data back to pressuring them to keep it from being released. The encryption was never the point. The leverage was.

Look at what that leverage has been pointed at. When Lehigh Valley Health Network was hit in 2023, attackers posted clinical photographs of cancer patients. More than 600 patients had medical record images published; the case settled for 65 million dollars. When Fred Hutchinson Cancer Center was breached that November, attackers emailed patients directly with their own data, and the centre confirmed it had notified the FBI and Seattle police about swatting threats made against those patients. When Medibank refused to pay, records belonging to 9.7 million Australians went up, including a file the attackers themselves labelled abortions.csv.

None of that required a technical advance. It required a decision.

Sextortion is the same decision, aimed lower. NCMEC now receives an average of 137 financial sextortion reports every day, a 37 percent increase on the year before, and says it is aware of at least three dozen teenage boys in the United States who have taken their lives as a result of being victimised by this crime. FinCEN's guidance to banks contains the detail I have not been able to put down: payments from minors may be between 10 and 50 dollars, while adult victims typically pay between 500 and 2,500. They are not doing this to children for the money. They are doing it because a frightened fourteen year old is easy, and the fourteen year old is the product.

What Erin West made me look at

I have been following Erin West's work for a while, and it has changed what I think this industry is actually up against.

She spent 26 years as a deputy district attorney in Santa Clara County and now leads Operation Shamrock, which she founded. She has been inside the compounds. Her congressional testimony this year describes a facility in Phnom Penh's Chinatown behind two story cement walls topped with razor wire, single controlled access gates, bars on the windows, eight people to a room. At an abandoned compound near the Vietnamese border she found rooms whose metal doors bolted from the outside, and fine schedules still posted on the walls: 300 dollars for touching the air conditioner, 100 dollars for misspelling a word.

She named the people. Shakilu, a Ugandan national, answered a Telegram job listing offering 1,500 dollars a month for a salesman role. He was taken to a torture room and beaten roughly 80 times. When he was no longer profitable he was sold to a second compound for 3,600 dollars, which the new operators told him he would have to work off.

The UN human rights office reported in February that credible estimates put the scam workforce at at least 300,000 people originating from 66 countries, and the industry in the Mekong region alone at more than 43.8 billion dollars a year. In October the US Justice Department seized approximately 127,271 Bitcoin, worth about 15 billion dollars, from the founder and chairman of Cambodia's Prince Group. The largest forfeiture action in the department's history.

I read This Is How They Tell Me the World Ends when it came out and found it frightening in exactly the way it intended, a book about an arms race and the exploits we bought and could not contain. Reading it against 2026, the frightening part has moved. The weapons are not what escalated. The labour model did. Somebody wrote a fine schedule. Somebody printed it and put it on a wall.

That is not a hacking problem. That is an industry, with an HR function, a supply chain and a recruitment pipeline reaching into 66 countries.

One honest Caveat, and The Number that Changed my Mind

I could make this louder than the evidence supports, and I would rather be accurate.

Phishing volume is not exploding. The Anti-Phishing Working Group observed 3.8 million phishing attacks in 2025, up slightly from 3.76 million in 2024. And the claims you keep seeing that AI now writes some specific majority of phishing email trace back to vendor marketing with no disclosed methodology and I am saying this as a vendor. 

But volume was always the wrong instrument, and the right one has finally shown up in the peer-reviewed literature rather than in a webinar.

At USENIX Security this month, researchers from TU Berlin, Inria and Ruhr-Universität Bochum published a field experiment with 7,700 participants. They gathered public information about each target, had a language model write an email tailored to that person, and measured who clicked. Personalised AI phishing produced a 10 percent click rate. Generic AI email produced 3.7 percent. Generic email written by a human produced 4.1 percent. Personalisation nearly tripled the click rate against both.

The cost of that personalisation was about three cents an email. Three cents. That is the number I would put in front of a board.

Now set it beside the FBI's ledger. Phishing and spoofing complaints in 2025 came to 191,561, down about one percent on the year. Reported losses on those complaints went from 70 million dollars to 216 million. Up 208 percent. Roughly the same number of attempts, three times the money out the door on each one.

A Harvard Kennedy School team put the number higher still, a 54 percent click-through for AI-written spear phishing against 12 percent for generic email, on par with human experts. It is the more quotable result and I am using the Berlin one instead, for two reasons. It was run in the field on 7,700 real addresses rather than on a hundred recruited volunteers. And it found something the louder version leaves out, which is that a human being writing by hand still beat the machine, 24 percent to 10.

That is worth sitting with, because it tells you what actually changed. The machine is not better than a skilled human attacker. It costs three cents, it never gets tired, and it will do this ten thousand times before lunch. What used to be reserved for the CFO is now affordable against the accounts payable clerk, the new hire, the person in your warehouse.

The volume of what arrives at the front door has not changed much. What arrives has. It knows your vendors, your approval chain and your CFO's turn of phrase, and it costs three cents to write. The door is easier to walk through than it has ever been.

We (the collective We that is) do still win things. Interpol's First Light operation produced 5,811 arrests across 97 countries this spring and intercepted 293 million dollars.

Who the mail in your inbox is actually after

Here is the finding that reorganised my thinking, and I should say plainly that it is ours, from our own platform, so weigh it accordingly.

In the second quarter of this year we looked at 32,000 emails that got past both the gateway and the behavioural tools and landed in a corporate inbox. Then we asked one question of each: whose money is this after.

Sixty eight percent were after the employee's. Not the company's. Consumer fraud, aimed at the individual, arriving on corporate infrastructure for the simple reason that is where the individual is. Another 23 percent were credential theft, which is the same thing one step earlier, a personal compromise that becomes a business problem later. Business email compromise, the category this entire industry organises its loss models and its product categories around, is the small slice left over.

Read that against everything above. Extortion moved from the company's servers to a patient's photographs. Sextortion never went near a company at all. Trafficked workers in a compound are not phoning your accounts payable department, they are phoning your employee's mother. And the mail landing on corporate mail systems is, by a margin of more than two to one, hunting the person rather than the organisation.

So it is not unwinnable. It is unwinnable on the current scoreboard.

Final Thoughts: The war we are actually in

We built an industry that is very good at measuring whether an attack was technically prevented, and almost incapable of measuring whether a person was harmed. Those used to be roughly the same question. The gap between them is where the last three years of loss growth lives.

So the honest version of my complaint is not that nothing changed. It is that we kept getting better at the game we knew how to score while the other side quietly changed sports.

If you are setting a 2027 plan, I would ask three things about it. Not what your detection rate is, because every vendor's number is good and none of them survives contact with your actual traffic. Ask instead where in this business coercion becomes possible, and who is holding the data that makes it possible. Ask what happens to the person on the other end when we get one wrong, not only what happens to the ticket queue. And ask whether anything you are funding this year makes it harder to run a compound, or only harder to reach you specifically.

That last one is uncomfortable, because most of what we all sell, mine included, answers the second half and not the first.

What I will say for our own corner of it is that two things have turned out to matter more than I expected when I got here.

The first is that the surface of these attacks is infinite and the structure underneath is not. Those 32,000 emails were reduced to 459 distinct designs, and twelve of those designs accounted for 78 percent of the volume. Signatures chase the surface, which is endless. Behavioural baselines chase what is unusual, and none of this is unusual any more. Reasoning about what a message is trying to do to somebody is the only one of the three that gets at the part small enough to count.

The second is explainability, and not for the procurement reason. A system that can say what an attacker was trying to do to a specific named person, in language that person would recognise, is the only kind of system that produces evidence. Evidence is what moves this from an IT expense to a prosecution. Erin West's argument, underneath all of it, is that the sector sitting on the evidence has not yet decided it is a participant.

I have been at this long enough to have watched a lot of things change. This is the one I want to still be arguing about when it does.

Frequently Asked Questions (FAQs)

Q1: Are cyberattacks actually becoming more sophisticated?

Not necessarily. The article argues that the bigger change is the increasing willingness of attackers to use familiar capabilities to cause greater harm. AI has made personalized attacks cheaper and easier to scale, but the underlying tactics are often familiar.

Q2: Why are cybercrime losses increasing even as security technology improves?

Detection, response, and information sharing have all improved, yet reported losses have continued to rise. FBI IC3 data cited shows reported internet-crime losses increasing from $12.5 billion in 2023 to $16.6 billion in 2024 and $20.877 billion in 2025.

Q3: How is AI changing phishing attacks?

AI makes personalized phishing much cheaper to produce at scale. The research cited in the article found a 10% click rate for personalized AI-generated phishing, compared with 3.7% for generic AI-generated email and 4.1% for generic human-written email.

Q4: Is phishing volume exploding because of AI?

The Anti-Phishing Working Group recorded 3.8 million phishing attacks in 2025, compared with 3.76 million in 2024. The more important change is the effectiveness and potential impact of increasingly personalized attacks.

Q5: Why are attackers increasingly targeting individuals?

Attackers are increasingly using personal harm as leverage. Examples include the use of stolen medical information for extortion, direct threats against breach victims, and financial sextortion. The objective is not always to disrupt a company's systems; it can be to pressure a person.

Q6: What did StrongestLayer find in its analysis of 32,000 emails?

In StrongestLayer's Q2 2026 platform analysis, 32,000 emails that had passed existing controls and reached corporate inboxes were examined. Sixty-eight percent were targeting the employee's money, while another 23% involved credential theft. The analysis also reduced those emails to 459 distinct designs, with 12 designs accounting for 78% of the volume.

Q7: What should CISOs consider when planning their 2027 security strategy?

Looking beyond traditional detection metrics and asking where coercion becomes possible, who controls the information that enables it. What happens to the person affected when security controls fail, and whether security investments reduce the ability to cause harm rather than simply reduce access to the organization.

‍