Someone cloned a CEO’s voice and stole €95 million with a phone call.

Thomson Reuters confirmed last week that fraudsters used AI to impersonate the CEO of Italy’s largest bank and stole €95 million from its private banking arm. One of the biggest deepfake scams in history.

How synthetic audio generation bypassed traditional executive authorization protocols.

Here’s how it worked.

In February the chairman of Fideuram—the private banking division of Intesa Sanpaolo—received an email that looked like it came from a well-known international law firm. Followed by a Microsoft Teams invite. Standard corporate communication. Nothing that would set off alarm bells.

Then came the phone call. A voice he recognized. A senior partner at that firm. And then another call—from someone who sounded identical to Carlo Messina, the CEO of Intesa Sanpaolo.

The voice was fake. The law firm partner was fake. The entire “confidential business opportunity” was fake.

But it sounded exactly right. The tone. The pacing. The way Messina spoke. The fraudsters didn’t hack a network. They didn’t exploit a firewall. They hacked a person’s ears.

The chairman authorized the transfers. €95 million moved to accounts in Hong Kong. By the time anyone realized what happened, the money was gone.

The chairman resigned in March. He’s not under investigation. He wasn’t in on it. He was just a guy who answered the phone, heard someone he trusted, and followed instructions.

So what do you do?

Stop approving anything important based on a phone call alone. Especially if it involves money. Wire transfers. Account changes. Access requests. None of it moves without a secondary out-of-band verification. A callback to a known internal extension. A verbal passphrase. An in-person confirmation.

Every company is adding AI right now. Almost none of them have tested what happens when it goes wrong.

ISACA just dropped their 2026 State of Cyber report. Surveyed over 1,800 cybersecurity professionals across the globe. One number stopped me cold.

Three percent.

Meanwhile 37% of companies are already using AI to automate security operations. 32% are using it for active incident detection. Tools making decisions about what gets blocked, what gets flagged, and what gets escalated.

So let me make sure this math is clear. Companies are handing real security decisions to algorithms they don’t fully control—and almost three-quarters of them have never once practiced what they’d do if that system failed. Hallucinated. Got poisoned. Or started leaking confidential data.

It gets worse. 56% of respondents don’t know how long it would take to detect an attack on their AI models. 39% don’t know if they even have a documented process for shutting an AI system down during an active breach.

“We’re flying blind with our hands on the throttle.”

Here’s what this looks like in the real world:

• An employee pastes client financials into ChatGPT. The data is now part of a training corpus outside your control.
• Your AI-powered email filter flags a legitimate client message as malicious and drops it—costing you a contract.
• An attacker uses AI-generated deepfake audio to impersonate your CFO on a vendor payment call.

If your answer to any of those is “we’d figure it out”… that’s not a response plan. That’s panic with a calendar invite.

You don’t wait for a fire to figure out where the exits are. And you don’t wait for an AI incident to ask who has the kill switch.

There’s an empty chair in every meeting I take.

Nobody sits in it but somebody’s always there.

I stole this idea from Amazon years ago. They put an empty chair in meetings to represent the customer. The idea was simple: whatever you’re discussing, the person paying the bills should be in the room.

I liked the concept but I changed it.

“My empty chair isn’t my best client. It’s my worst one.”

The empty chair that represents your most demanding, skeptical client.

The one who questions every invoice. The one who notices when a cable isn’t labeled. The one who calls at 6:45pm on a Friday because an access badge reader took two seconds longer than usual.

That’s who I imagine sitting in the room when we’re about to make a decision.

Because here’s what I’ve learned about the difficult clients over two decades in this business:

They’re not difficult because they’re unreasonable. Most of the time they’re difficult because their standards are higher than everyone else’s. And they don’t give you credit for doing 90% of the job right when the last 10% is where things break.

Your easy clients will say “sounds great” to almost anything. They won’t push you. They won’t spot the holes in your logic. They’ll let you get comfortable.

The hard client sitting in that empty chair makes you better. Every single time.

Last month we were about to roll out a new internal process. Team was excited. Everyone agreed it made sense.

Then I looked at the chair. And I thought, “Mike would tear this apart in thirty seconds.”

So I asked the room—“what would our toughest client say about step three?”

Silence. Then someone said: “They’d ask why step three exists if step two already did the job.”

We cut step three. The process got simpler. The rollout went smooth.

The cameras watching your building are watching you. But they might also be watching for someone else.

Between June 17th and July 22nd a single operator scanned the internet and compromised more than 14,500 Dahua-branded cameras and digital video recorders across dozens of countries.

Here’s how simple it was. The attacker ran three methods at scale:

That backdoor account survives password changes. On most firmware versions, a factory reset doesn’t even clear it.

Let that land. You could change every password on your camera system today, feel good about your security, and the attacker still has admin access to your live video feeds.

And how did researchers find all of this? The attacker left their command-and-control server exposed without authentication. Security researchers literally walked in and downloaded the whole database.

“The hacker got hacked by their own laziness.”

Here’s why this matters beyond the headlines. A compromised camera isn’t just someone watching your loading dock. It’s an unmonitored computer connected directly to your internal network. It can sniff traffic. It can bridge network segments. It can be used as a persistent pivot point to reach everything else.

If you have security cameras in your building—and most businesses do—ask your IT or physical security team three questions:

Protecting business facilities requires hardened, verified hardware. Learn how Smart Pro installs enterprise-grade, NDAA-compliant commercial surveillance and security camera systems across the Kansas City metro area.