One firm lost US$25 million to a video call full of colleagues who weren't real.
The full script, word for word — 524 words, about 2 minutes to read. Current as at July 2026. AI tools change quickly; if something looks different when you try it, check the product's own help pages.
A finance worker at the engineering firm Arup joined a video call with his chief financial officer and several colleagues he recognised. He'd been suspicious of the email that set it up — it asked for a confidential transaction, and it smelled wrong. The video call is what convinced him. However, every other person on that call was an AI-generated fake. He made fifteen transfers that day, totalling about twenty-five million US dollars. None of it has been recovered.
No system was breached. Nobody on the inside helped. The attackers built those fakes from footage that was already public — recorded webinars, press interviews, conference sessions sitting on YouTube. Which means the exposure isn't a technology problem you can buy your way out of. If your finance director has ever spoken at a conference, or your CEO has done a podcast, the training material exists. For most organisations that's a list of the same handful of people who can also authorise a payment.
Australia gives a sense of the scale. Two point one eight billion dollars lost to scams there in twenty twenty-five — and while the number of reports has levelled off, losses went up nearly eight per cent. Investment scams take the largest share by a distance, and they're the category where AI-faked endorsements do the most work. One Australian woman lost her life savings after watching a deepfaked video of a well-known businessman and clicking through.
It turns up in three shapes. Someone you report to, asking for something urgent and slightly outside process. Someone well known, apparently endorsing an investment. Or someone who's been messaging you warmly for weeks and now has a problem that money would solve. All three existed long before AI. What changed is that the voice on the phone now sounds right, and the face on the call now moves right — so the two checks people used to rely on have quietly stopped working.
The defence is procedural rather than perceptual. If a request involves money or credentials, verify it on a channel the other person didn't choose — end the call, and ring the number you already had for them. At home, agree a spoken phrase with your family that a stranger couldn't know, and use it if someone rings in distress. And at work, make it a standing rule that no payment goes out on a single approval, whatever the seniority of the person asking. Rules survive pressure. Judgement, at four-thirty on a Friday, does not.
Looking for the tell — the strange hands, the blinking that's slightly off, the mouth that doesn't quite sync. That worked for a while. It doesn't now, and every month it works less. Any defence built on spotting the flaw has a short shelf life. Build yours on how you verify instead.
One thing to do this week. Pick the person most likely to be impersonated to you — your boss, your accountant, your mum — and agree with them right now how you'd check it was really them. Do it before you need it, because in the moment you need it, you won't be thinking clearly.