The next voice that robs your bank will sound exactly like someone you trust. An SA company has moved first to stop it

The next voice that robs your bank will sound exactly like someone you trust. A South African company has moved first to stop it.

---

South African AI company Untapped AI has filed two patent applications with the United States Patent and Trademark Office (USPTO) covering technology that authenticates the voice on a phone call while the call is happening and cuts the call when that voice is not what it claims to be. The filings, lodged in September, go after the problem now keeping banks, insurers, telecoms operators and governments awake at night. When any voice can be copied in seconds, nobody can trust a voice on the phone, and Untapped AI has built the answer.

The latest crime figures show how exposed this country already is. South Africans lost R2.4 billion to digital banking crime in 2025 across more than 110 000 reported incidents, according to the South African Banking Risk Information Centre's (SABRIC) 2025 Annual Crime Statistics Report. Most of those crimes were driven by social engineering and the manipulation of customers rather than attacks on bank systems. A criminal did not need to break into a bank. A criminal only needed to make a convincing phone call.

Those losses were recorded before AI voice fraud reached this country at scale. Voice cloning tools are now freely available, and a single AI system can run thousands of scam calls at the same time, each hidden behind a spoofed number that looks exactly like a bank, an insurer, a medical scheme or a government department. Untapped AI expects a sharp rise in fraudulent calling and voice impersonation across South Africa in the months ahead as these tools move into the hands of organised crime.

South Africa's defences have not kept pace. Most South African call centres today have no way of knowing whether the customer on the line is a clone. Most finance teams cannot tell whether the CEO approving an urgent payment is the CEO at all, and most customers have no way of confirming that a call from their bank came from their bank.

The criminals already know where the gaps are, and the first South African business to have its name attached to a cloned call in the news will learn it the hard way.

In February this year, fraudsters took €95 million, roughly R1.9 billion, from Fideuram, the private banking arm of Italy's largest lender Intesa Sanpaolo. According to Reuters, the scheme began with a WhatsApp message that appeared to come from the group's chief executive and was sealed by a phone call in which criminals used AI to clone the voice of a senior partner at a prominent law firm. The bank's then chairman believed what he heard and instructed a series of transfers to accounts mainly in China and Hong Kong. About €36 million has never been recovered after being moved through overseas accounts and converted into crypto-currency.

Fideuram had the resources and security teams of one of Europe's biggest banking groups behind it, and a single cloned voice still walked straight through. It is far from the only case.

In 2024, an employee in the Hong Kong office of global engineering firm Arup joined a video call with the company's chief financial officer and several colleagues, then made 15 transfers worth US$25 million on their instructions. Every other person on that call was an AI-generated fake. In 2025, an impostor used an AI clone of US Secretary of State Marco Rubio's voice to leave messages for foreign ministers and senior American officials, and in the same year, fraudsters imitated an Italian government minister to talk businessman Massimo Moratti into sending close to €1 million to an overseas account.

The attacks have also become personal. In the United States, Arizona mother Jennifer DeStefano answered her phone to hear her 15-year-old daughter sobbing and begging for help, followed by a man demanding a ransom. Her daughter was safe on a ski trip the whole time and the voice on the line was an AI clone. DeStefano later testified about the call before the US Senate. Anyone who has ever posted a video or left a voice note online has already handed criminals enough audio to do the same to their family.

The cost of running these attacks has collapsed. Dr Matthew Canham, a former FBI agent, has said that most voices can be cloned from between 15 and 45 seconds of audio, and that AI agents have been capable of carrying out these attacks on their own, with no human criminal on the call, since around 2023.

Untapped AI's two patent filings change where trust sits on a call. Today, a caller proves who they are once, through a PIN or a set of security questions, and everything said after that is taken on faith. Untapped AI verifies the voice continuously for the full duration of the call and identifies where that voice is really coming from, not just what it sounds like.

The first filing covers the authentication itself. Proprietary models recognise signatures only Untapped AI can read, and those models change continuously on the back end, so there is never anything fixed for a criminal to copy or replace. Every authenticated call carries a verification stamp, comparable to the security seals consumers already rely on when they pay online. A call that fails verification, or that someone tries to tamper with, is terminated automatically before any damage is done.

The second filing extends that protection to the AI agents now making and receiving calls on behalf of businesses. As companies deploy voice agents for sales, collections and customer service, the question of whether the agent on the line is the legitimate one, or a criminal's copy, becomes as pressing as it is for a human voice. Untapped AI's technology gives both parties live confirmation that the person or AI agent speaking is who they claim to be.

The company is not disclosing how the technology works. The method behind it is protected both by the patent filings and by trade secrets that never leave the business.

The Fideuram call is exactly the moment the technology was built for. The cloned voice of the lawyer was what turned a suspicious message into an approved transfer. On a call protected by Untapped AI's authentication, that voice would have carried no verification stamp, and the system is designed to end such a call before a single instruction is acted on.

The same protection applies down the chain. A call centre agent would know within seconds that the customer asking to change a payment beneficiary is a clone. A finance clerk would never hear the fake CEO's instruction, because the call would already be over. A pensioner would know that a call from the bank came from the bank, without having to spot a scam on their own.

Banks and insurers invest heavily in know your customer (KYC) checks, and those checks do their job at onboarding by confirming who opened an account. They cannot confirm who is speaking on a call six months later, and that gap is exactly where fraudsters operate. Untapped AI's technology takes verification beyond KYC and into the conversation itself, closing the gap that a judder in a customer's voice or a wrong security answer was never going to close.

"Every bank in the country tells its customers that it will never phone and ask for their PIN, and every month criminals phone anyway and get it," said Lloyd Matthew, Director of Untapped AI. "Warnings have stopped working, so we built something that does not rely on the customer getting it right. When a call carries our authentication, the person on the other end knows it is real. When it does not, or when someone tries to tamper with it, the call ends.

"People ask me how it works, and the closest comparison I can give is Shazam," Matthew added. "Shazam hears a few seconds of a song in a noisy room and tells you exactly what you are listening to. We do something similar for the human voice, and we can tell you where that voice is really coming from. The difference is that our models never sit still. They keep changing behind the scenes, so by the time a criminal has learnt anything about them, there is nothing left to copy. I won't say more than that, and neither will anyone else in this company.

"Look at what happened in Italy," Matthew said. "One of the most sophisticated banking groups in Europe lost close to R2 billion because a voice on the phone sounded right. If it can happen there, it can happen at any bank, insurer, medical scheme or municipality in South Africa tomorrow. It matters to me that the answer comes out of South Africa, built on sovereign local infrastructure, for South African institutions first and then for every market facing the same threat."

Untapped AI is in discussions with government and major players in the South African telecommunications sector on how call authentication can become a national standard, so that every business has access to the same protection instead of fighting the threat alone. The company believes South African organisations should treat call authentication with the same seriousness they give POPIA compliance and cyber security certification, and that those who move first will set the level of trust their customers come to expect.

"We saw this threat coming before most of the market did, and we moved early to secure and build the technology to stop it," said Matthew. "That technology exists today. What South Africa needs now is an agreed standard, because fraud on this scale will never be beaten one company at a time."

The two filings form part of a wider portfolio of patent applications held by Untapped AI covering voice provenance and the authentication of both human and AI callers. Further filings will be announced in due course.