Key Takeaways
- Out-of-band verification means confirming a request on a second channel the attacker doesn’t control.
- Detection is dead. Arup’s employee was suspicious, asked for a video call to verify, and every face on the call was a deepfake. About $25.6 million gone.
- Ferrari survived on luck: one executive, one unanswerable question. Luck doesn’t scale. A procedure does.
- Pick your irreversible actions, write the one-line rule, and make questioning the message policy instead of insubordination.
In January 2024, a finance employee at the engineering firm Arup got a suspicious email about a confidential transaction. The employee did the right thing and pushed back. The criminals arranged a video call so the employee could verify the instruction directly with the people involved.
On the call: the CFO. Colleagues the employee recognized. Voices that matched. A coherent conversation. The employee authorized fifteen transfers totaling about $25.6 million.
Every person on that call except the employee was an AI-generated deepfake, built from publicly available footage.
The employee was not careless. The employee followed the old advice: be suspicious, verify before acting. The old advice failed because it never said where the verification has to happen.
The FBI’s Internet Crime Complaint Center put business email compromise losses at $3.046 billion in 2025. And 2025 was the first year the Bureau formally tracked AI-enabled fraud: $893 million across more than 22,000 complaints involving deepfaked audio and video. No malware. No hacking. Just messages that look and sound exactly like the real thing.
What It Is
Out-of-band verification is a simple idea with an ugly name. When a message asks you to do something irreversible, you confirm it on a different channel than the one the message arrived on.
The email says wire the money. You call the person, on the number already in your contacts, and ask. The text says the boss has a new number. You call the old number and ask. The vendor says their bank details changed. You call the vendor’s main line and ask.
Out of band means outside the attacker’s channel.
Luck Is Not a Policy
Six months after Arup, Ferrari faced the same script. A Ferrari executive got WhatsApp messages from the CEO’s number: a big acquisition was coming, sign the NDA quickly, utmost discretion. Then a phone call, in a voice nearly identical to the CEO’s, pressing a currency deal tied to the acquisition.
The executive asked one question the caller could not answer: name a book the real CEO had recently recommended to him. The line went dead. Ferrari lost nothing.
Brilliant. Also luck.
No procedure produced that question. One person happened to think of the exact thing a deepfake couldn’t fake, in the moment. You cannot hire that moment. You cannot train for it, because you never know the attack in advance. The next executive might not think of anything at all.
And strictly speaking, that wasn’t out-of-band verification. The executive never left the call. He tested the caller with a shared secret, which is a cousin of the real thing, not the thing itself.
Policy is the Ferrari outcome without the luck. Out-of-band verification doesn’t need anyone to think of the perfect question. It just needs them to make the call.
How It Works
Out-of-band verification has three moves.
Pause. Urgency is the attacker’s best tool. The sign-it-now NDA, the confidential transaction, the boss who needs it now. A real request survives a ten-minute delay. A fake one depends on it.
Switch channels. If the request came by email, verify by phone. If it came by text, verify by a call to a known number. If it came on a video call, verify through a different medium entirely. The new channel must be one you chose, not one the message handed you. Call the number in your contacts, not the number in the signature.
Confirm the action, not the story. You are not asking “did you send this email?” You are asking “are you asking me to approve this transfer today?” Confirm the action, because the action is what costs money.
Why It Is So Effective
Every one of these attacks has the same shape: the criminal owns exactly one channel. The compromised inbox. The spoofed number. The deepfaked video call. All of their power lives inside that channel.
The second channel breaks the monopoly. The attacker who can fake an email cannot also answer your phone call to the real number. The attacker who can deepfake a video call cannot intercept the callback to a number they have never seen.
Notice what this does not require. It does not require anyone to spot a fake. It does not require training, software, or a security team. It works against AI-generated fraud for the same reason a deadbolt works against a lockpick: it never plays the attacker’s game. It changes the game.
And it is nearly free. The most effective control in the building is a phone call.
Do It Today
Pick your irreversible actions. For most organizations there are three: money moving, payment details changing, and access being granted or reset. Password resets and new logins belong on this list too.
Write the rule in one line: nothing on this list happens on the word of a single message. Ever.
Then tell everyone, and mean the hard part: questioning the boss’s email is not insubordination. It is the policy. A workplace where verifying an executive’s request feels like defiance has no verification policy at all. It has a hope policy.
Start with the next request that arrives. Pause, switch channels, confirm the action. A week from now you will wonder why this was ever optional.