The phishing email didn’t disappear.
It just got a ChatGPT account.
Microsoft recently reported that attackers are impersonating some of the biggest names in AI—ChatGPT, Microsoft Copilot, Claude, and DeepSeek—to get people to click, download, log in, or hand over payment information.
In one ChatGPT-themed campaign, as many as 100,000 emails were sent in a single day.
The message told people they needed to update their ChatGPT Plus payment information.
People clicked.
Some handed over personal information and credit card details.
And before anyone says, “Well, I’d never fall for that,” I’d be careful.
Because that’s exactly the attitude phishing relies on.
The technology changed. The psychology didn't.
I’ve been doing this long enough to see a lot of versions of the same scam.
There was the fake shipping notification.
The DocuSign request.
The Microsoft password reset.
The invoice that needed immediate payment.
The email from the CEO who was supposedly in a meeting and desperately needed you to buy gift cards.
The details change.
The underlying formula rarely does.
Borrow trust from something familiar.
Create just enough urgency or curiosity.
Get someone to act before they stop and think.
AI is simply the newest—and currently very effective—costume.
That matters because AI tools are quickly becoming normal parts of the workday.
Employees are signing up for tools, testing new platforms, receiving product updates, joining trials, installing plugins, and being asked to experiment with AI.
So an email that says your AI account needs attention doesn’t necessarily feel unusual anymore.
It feels like work.
And that’s exactly what makes it useful to an attacker.
We're also making the fakes better.
There’s another side to this.
AI isn’t only something attackers can impersonate. It can also help them create more convincing material.
Microsoft recently documented a separate campaign involving more than a million emails targeting enterprise users.
The attackers impersonated company executives and a legitimate vendor, created fabricated invoices, constructed fake forwarded email conversations, and attempted to convince accounts-payable teams to send ACH payments of nearly $50,000.
Microsoft found indicators consistent with generative AI being used in the development of the email templates.
Think about what that means for the person receiving the email.
The old advice was to look for spelling mistakes, strange formatting, awkward language, or something that just felt “off.”
Those clues aren’t gone.
But they’re becoming less reliable.
An attacker no longer needs to write a convincing email from scratch.
The fake invoice can look professional.
The email thread can sound plausible.
The CEO can sound like the CEO.
The vendor can look like the vendor.
The attackers don’t need everything to be perfect.
They just need one person, on one busy Tuesday afternoon, to believe it for thirty seconds.
Stop training people to spot bad emails.
This is where I think companies need to rethink some of their security training.
We’ve spent years teaching people to recognize suspicious emails.
That’s still useful.
But recognition can’t be the entire defense.
Because what happens when the suspicious email doesn’t look suspicious?
You need processes that don’t depend entirely on someone’s ability to detect a fake.
A request to change payment information should be independently verified.
An unusual financial request from an executive should have a second approval path.
A login request should happen through the application you already use—not through the link that conveniently arrived in your inbox.
Unexpected software shouldn’t be installed because an ad or email says there’s a shiny new AI tool everyone needs.
And people need permission to slow things down.
That’s important.
A lot of social engineering works because organizations reward speed.
Respond quickly.
Get the payment out.
Handle the request.
Don’t bother the boss.
Attackers understand that culture remarkably well.
Sometimes the most security-conscious thing an employee can do is become mildly inconvenient.
Call the person.
Open the service directly.
Ask the annoying question.
Make someone verify the request.
I’d rather lose five minutes than $50,000.
Technology still matters.
None of this means we throw our hands up and make employees responsible for stopping every attack.
Quite the opposite.
Humans should be one layer of the defense, not the entire defense.
Email authentication, anti-phishing protections, identity controls, safe-link scanning, endpoint protection, multifactor authentication, monitoring, and properly configured security tools all matter.
So do boring things like reviewing who has access to what and making sure your payment processes aren’t built entirely around trust.
There is no single product that solves this.
There never was.
Security works best when the technology and the human process assume that eventually somebody is going to click something they shouldn’t.
Because eventually somebody will.
I’ve been in technology for nearly three decades.
People haven’t suddenly become less intelligent.
The environment around them has become much more convincing.
That’s a different problem.
The Forecast
The weather’s getting rougher out there.
But not necessarily because every cyberattack is becoming radically more sophisticated.
Some are.
Others are simply becoming better at looking normal.
That’s the shift leaders need to understand.
The next generation of phishing may not arrive with broken English, a bizarre attachment, and a Nigerian prince.
It may arrive looking exactly like another tool your company started using six months ago.
So don’t build a security culture around spotting yesterday’s scam.
Build one around verification.
Trust—but verify—is old advice.
Turns out it aged pretty well.
