How AI voice clones fake a kidnapping call, and the one word that stops it

According to ESET, AI voice scams no longer need much to work. A few seconds of audio, scraped from a social media post, is enough to clone a voice convincingly. The technology is cheap, it is getting better fast, and it has lowered the barrier for anyone who wants to try this.

ESET describes one of the most common versions of the scam: a cold call to a relative, playing a deepfake voice designed to sound like a loved one who has been kidnapped. The scammer adds personal details pulled from the same social accounts, so the call feels specific rather than generic. Sometimes the scammer watches the victim's own social media, or a relative's, to pick the best moment to call, ideally when that person is travelling and less likely to check in with family directly.

Why the call sounds so real

The voice clip itself is short. ESET notes that scammers use the cloned voice for only a few seconds at a time, mixed with sobbing, sometimes layered with background noise. That is not an accident. A short clip is harder to scrutinise than a long one, and distress plus noise gives a listener less to focus on. The goal is to keep the relative frightened and guessing, not to give them time to think clearly.

The scam depends on one thing above all: the victim trusting their own ear. Recognising a family member's voice has always been treated as proof of identity, something no stranger could fake. Cloning technology breaks that assumption quietly, without the victim ever realising the basis for their trust has changed.

The fix is a word, not a filter

ESET's suggested defence does not rely on spotting the fake. It relies on a detail no scraper can find. A family agrees in advance on a safe word or phrase, something a relative can ask for if a call sounds off. A scammer working from scraped audio and social posts has no way to produce that word, because it was never posted anywhere public.

The word has to be chosen carefully. ESET warns against anything guessable from open-source information: a pet's name, a favourite sports team, anything that shows up in a photo caption or a bio. The whole point is that the word exists nowhere a scammer could scrape it.

This shifts the test away from something the scam is built to defeat. The scammer's method banks on the voice sounding right and the moment feeling urgent. A safe word does not ask whether the voice sounds right. It asks for something the voice, however convincing, cannot supply on its own.

ESET also flags the human side of setting this up: families need to hear why the precaution matters, in a way that explains the technology has moved on, without frightening anyone into thinking every call is now a threat.