• A familiar voice or face is no longer proof of identity.
• Attackers build their fakes from public audio and video.
• Urgency plus secrecy is the scammer’s signature.

Last time, in The Unfriendly Skies, Sarah dodged three Wi-Fi traps on a business trip, each one built to make something fake feel familiar. This time, the fake is a voice you know, or a face on a video call. Welcome to AI and deepfakes.

It’s 4:47 p.m. on a Thursday, and you are one email away from freedom.

Your phone rings. The screen says it’s your CEO.

You answer, and the voice is unmistakable: the cadence, the slightly weary patience, the particular pause before asking for something inconvenient.

“I need your help with something, and it needs to stay between us.” There’s an acquisition. A payment has to go out today. The usual approvers can’t know yet. Timing is everything. Very sensitive.

You hesitate. The line gets a few degrees colder.

“Are you seriously telling me you don’t recognize my voice?”

Here’s the plot twist: you do recognize it. That’s the problem. And the right answer is: “Yes, I do. I’m going to call you right back anyway.”

Because the voice on the line may have been built by a criminal from a conference talk, a podcast episode, and some widely available software. This is vishing, short for voice phishing: a scam run over the phone, now with AI-powered impersonation as an add-on.

It’s also not a fringe problem. Mandiant, Google’s incident-response team, reviewed what attackers did to get into companies in 2025. Voice phishing was the second most common way in, behind only exploited software flaws. It showed up in 11% of investigations where the entry point could be identified, nearly double email phishing at 6%. For a decade we trained everyone to side-eye their inbox. The criminals noticed and picked up the phone.

Here’s what’s ahead: how the trick is built, six true stories (some ended in tears, some in triumph), the tells that actually matter, six habits that beat it, and a straight answer to whether we should refuse to let vendors record our video calls. Short version of the whole article: we’ve got this. Plenty of people in these stories got it right.

Your voice is basically open source

For most of human history, recognizing a voice was a perfectly good way to know who was talking. Accent, rhythm, favorite phrases: hard to fake, easy to recognize. AI has made them reproducible.

Voice cloning is software that studies recordings of a person and generates brand-new speech in that voice. The legitimate uses are real: audiobook narration, dubbing videos into other languages, restoring speech for people who have lost it. The same tools in other hands can manufacture a payment instruction from your CEO, a plea from your child, or a “security check” from someone claiming to be IT.

Where does the raw material come from? Everywhere you have ever spoken in public: earnings calls, conference panels, podcasts, webinar replays, LinkedIn videos, YouTube, a chatty voicemail greeting. Consumer-protection researchers report that a few seconds of clear audio can be enough for a passable copy, and more audio makes a better one. When Consumer Reports tested six popular voice-cloning products in 2025, four had no meaningful safeguards beyond asking users to promise they had permission to clone the voice. The person being cloned never has to take part, or even find out.

Voice cloning is one flavor of the deepfake: audio, video, or images created or altered by AI to make someone appear to say or do something they never did. You’ll meet the video version in a moment.

How a vishing attack is built

It looks like improvisation. It’s closer to a recipe, and professionals follow it.

1. Homework. LinkedIn, the company website, press releases, conference agendas. Attackers learn who reports to whom, who can approve payments, who works in finance or at the help desk, and what’s in the news. (A real acquisition makes a fake one far more believable.) It’s tedious, and almost all of it is public.

2. The voice. The premium option is a clone of a real executive’s voice, used as pre-recorded messages or converted live as the attacker talks. The budget option needs no AI at all: a confident caller who knows names, systems, and internal jargon. Both work. One of the most expensive breaches in recent memory used the budget option (more on that below).

3. The disguise. Spoofing means faking the caller ID, and it’s as easy as it sounds: caller ID is a label the caller supplies, not a verified fact. Or the attacker opens a brand-new WhatsApp, Signal, or Teams account using the executive’s public photo. Often a text arrives first to warm you up and move the conversation somewhere with less oversight.

4. The pressure. Urgency, secrecy, authority, and a hint of disappointment. We’ll look at why this works in a minute.

5. The ask. It’s almost always one of a short list: send a wire or change a vendor’s bank details; read out or approve a login code; reset a password or MFA (multi-factor authentication, the extra step like a code or phone prompt that proves it’s really you) for “an employee who’s locked out”; install a “support tool” that hands over remote control; or visit a “verification” page that is actually a fake.

6. The cash-out. Stolen money is split across accounts and moved fast. Stolen access is often handed off to another criminal crew for ransomware or data-theft extortion; Mandiant reports these hand-offs are increasingly routine.

Six true stories

All documented, all real. Pay attention to what the people who won did differently.

Case file one: The melody (UK energy company, 2019)

The CEO of a UK energy firm took a call from his boss, the chief executive of the German parent company. The accent was right. So was the melody of the speech. The boss wanted €220,000 (about $243,000) sent to a Hungarian supplier within the hour, with a promise of quick reimbursement. He paid. A second call claimed the money had been repaid. A third asked for more. When the reimbursement never showed up and the third call came from an unfamiliar foreign number, he got suspicious and stopped. The first payment, meanwhile, had already hopped from Hungary to Mexico and onward. The company’s insurer covered the loss and shared the story, widely reported as the first known AI voice fraud.

The takeaway: that was 2019. The tools have improved enormously since; the script hasn’t changed a bit. The voice was convincing, but the story was the tell: urgent, unusual, a supplier he didn’t know, and a refund that was “on its way.”

Case file two: The fake boardroom (Arup, 2024)

In January 2024, a finance employee in the Hong Kong office of the engineering firm Arup received an email, apparently from the company’s CFO in the UK, about a confidential transaction. Smart move: he suspected phishing. Then came the fix the attackers had prepared. He was invited to a video call with the CFO and several colleagues, all of whom looked and sounded right. Hong Kong police later said that everyone else on the call was a deepfake, built from publicly available video and audio. Reassured, he made 15 transfers totaling HK$200 million (about $25.6 million) to five bank accounts. The fraud came to light afterward.

Arup later confirmed the incident and said its systems were not compromised. Its global CIO said publicly that attacks of this kind, including invoice fraud, phishing, WhatsApp voice spoofing, and deepfakes, were rising sharply, and he shared the story hoping it would help others.

The takeaway: the employee’s instinct was right. The attackers simply solved his doubt for him, using their own channel. Video is no longer proof of identity, and a check that uses the attacker’s own channel is not a check. The rescue here was never going to be a sharper eye; it would have been a second, independent confirmation, like a call to a known number, or a second approver.

Case file three: The book report (Ferrari, 2024)

In July 2024, a Ferrari executive got a string of WhatsApp messages from “CEO Benedetto Vigna.” They came from a number that wasn’t Vigna’s usual business mobile, with a different profile photo, and they hinted at a big, secret acquisition, with an NDA on the way. Then the phone rang. According to Bloomberg, the voice was an impressive imitation of Vigna’s southern Italian accent.

The executive got a bad feeling and tried something wonderfully low-tech. “Sorry, Benedetto, but I need to identify you.” What was the title of the book Vigna had recommended to him a few days earlier? The caller couldn’t answer, and the call ended. Ferrari opened an internal investigation and lost nothing.

The takeaway: notice how many warning signs arrived before the voice: a new number, a changed photo, an urgent secret deal. And notice the winning move: refusing to treat a convincing voice as proof. A personal question works as long as the answer isn’t findable online, which is why a pre-agreed code word is even better.

Case file four: The meeting with no CEO (WPP, 2024)

Fraudsters set up a WhatsApp account using a public photo of WPP CEO Mark Read, then used it to arrange a Microsoft Teams meeting with another senior executive and an agency leader. In the meeting they used a voice clone and YouTube footage of the other executive, while “Read” took part only through the chat window. The pitch: help set up a new business, which would be a vehicle for extracting money and personal details.

It failed. The people involved smelled something wrong, and Read then emailed the whole leadership team with the details and a line worth printing on a mug: “Just because the account has my photo doesn’t mean it’s me.”

The takeaway: this is what winning looks like, and it has two halves. The target noticed and stopped. Then the company told everyone, so the next target had a head start.

Case file five: Hello, IT? (MGM Resorts, 2023, and the help-desk wave since)

In September 2023, attackers reportedly found an MGM employee’s details online, phoned the company’s IT help desk posing as that employee, and talked their way into access. The fallout reached casino floors, hotel systems, and websites, which were disrupted for days. MGM later estimated the damage at around $100 million, and a rival casino company hit by the same crew reportedly paid roughly $15 million in ransom. No AI voice was needed. Just a phone, preparation, and confidence.

Since then, the play has been industrialized. A caller posing as IT, the help desk, or HR tells an employee something needs “verifying,” steers them to a lookalike company login page, and collects the password and MFA code in real time. Google’s threat-intelligence team documented a 2025 campaign that used voice calls to talk employees at many organizations into authorizing a malicious connected app in their company’s Salesforce account. Mandiant’s 2026 report found voice phishing was the single most common way into cloud environments in 2025, at 23%. Voice-cloning tools make these calls more convincing, but they aren’t required.

The takeaway: the target isn’t always the executive. The help desk, the new hire, and anyone with a login are all doors. And “IT is calling” is just as spoofable as “the CEO is calling.” Case file 6: It isn’t only business (the FBI warning, 2025)

Case file six: It Isn't only business (the FBI warning, 2025)

In May 2025, the FBI warned that since April, criminals had been sending text messages and AI-generated voice messages impersonating senior US officials, aiming to build rapport with current and former officials and their contacts and then get into their personal accounts. The warning also advised treating any message claiming to be from a senior official as unverified until proven otherwise. In mid-2025, a State Department cable reportedly described someone using AI to pose as Secretary of State Marco Rubio on Signal, contacting at least five people, including three foreign ministers.

The takeaway: if the technique is worth trying on foreign ministers, it’s worth trying on the rest of us. And the same tools power the “Mom, I’m in trouble” call to families, so everything in this article protects your home life too.

Every one of these attacks followed the same script, and so do the defenses. In Part Two: why smart people say yes, the tells that actually matter, six habits that beat it, and whether we should let vendors record our video calls.