Cape Town company Untapped AI says it filed two patent applications in the US in September for technology that can prove where a phone call is really coming from. It wants to stop criminals who use AI to clone voices and pose as banks, bosses and family members.

The technology has been tested but is not live. It will only protect people once mobile networks or businesses install it at both ends of a call.

Untapped AI builds AI voice agents for businesses and has offices in Cape Town and Johannesburg. Its backers include business telecoms and contact-centre provider Yaxxa, whose CEO, Heinz Vollmer, confirmed the investment.

Director Lloyd Matthew said the technology has two parts, and he would only describe them in broad terms, for security reasons. Both involve a cryptographic key that changes with every call. Matthew said the key is either swapped at the network switch when the call connects or carried inside the call audio as a signal people cannot hear.

The second part “analyses the sound’s frequency, gaps and other signal markers to verify the integrity of the conversation, rather than simply comparing a caller’s voice against a recording of someone’s voice”, Matthew said. “We verify where the voice is coming from, which means the source of the call.”

He compared it to a bank card. “Using a deepfake to get through our security is like a person expecting to draw cash from an ATM with a stolen card while wearing a mask of the owner. The machine doesn’t look at your face. It wants the Pin.”

Already proving costly

The company’s press release describes technology that authenticates “the voice on a phone call”, while Matthew’s written answers to TechCentral said the patents “don’t analyse the voice at all”. Asked about the difference, he said: “Both are accurate, and the wording could have been clearer.”

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 (Sabric).

Sabric told TechCentral it cannot say how much of that loss started with a phone call. It said AI voice cloning is an emerging risk, but “current evidence does not suggest that it has overtaken conventional social engineering scams in South Africa”.

Overseas, cloned voices are already proving costly. In February, fraudsters took €95-million from Italian private bank Fideuram, part of Intesa Sanpaolo, using a fake WhatsApp message followed by a phone call in which AI was used to clone a senior lawyer’s voice, Reuters reported. More than half of the money has since been recovered.

Asked whether the technology protects anyone before it is installed, Matthew said: “No. The protection has to verify on both sides of a connection, so it starts working once a network or business has deployed it at each end. That is why our first conversations are with the networks, because one deployment there covers every subscriber.” Untapped AI has entered discussions with a major mobile network, which he would not name.

Untapped AI would not explain its process in detail or publish test results. “We are deliberately vague beyond that, because we don’t want to describe the actual process publicly,” Matthew said.

Prof Martin Olivier of the University of Pretoria’s department of computer science said he could not fully judge the claim from the information available.

“It’s a well-known mantra of cryptography that secrecy tends to point to something about the system that needs to be hidden,” he said. “All the details of the cryptographic protocols that we trust, such as AES and RSA, are publicly available.” AES and RSA are widely used encryption standards.

He said a proven cipher, used properly, would stop a clone. But networks that compress audio “may distort signals”, he warned.

In the US, networks use a call authentication standard called Stir/Shaken, which signs the caller’s number rather than the audio. Olivier said it works much like the certificates that make websites trustworthy. Matthew said his system does not rely on the number shown on screen.

Andries Maritz, principal product manager at Entersekt, said putting the signal inside the call audio “is a meaningful step forward, with conditions”.

Not a standalone solution

“A signal inside the audio is tied to the call itself, so it can’t easily be separated from it,” Maritz said. But a network can only check calls that cross it. “A fraudster using VoIP or another route bypasses it, so banks still need other controls.” VoIP refers to calls made over the internet.

Sabric said call authentication could help close a gap on phone calls. “However, it should not be viewed as a standalone solution.” It warned that criminals adapt quickly and “may increasingly exploit alternative channels such as WhatsApp, SMS, e-mail or social media platforms”.

Banking technology executive George Chinomona, speaking in his personal capacity, said AI mostly makes old scams work better. “Banks should assume that voice alone can no longer be treated as proof of identity.”

Matthew wants government to change who carries the loss when a scam call works. “We hope government will make policy so that responsibility for call fraud moves from the customer to the business. If that happens, we will share our code with the market, competitors included.”

Loading ...

MTN said any national framework should be built jointly by industry, regulators and technology providers. It said open and interoperable standards “encourage broader adoption, innovation and consistency across networks”.

Until any of this is live, Matthew’s advice is simple. Agree on a family code word. If a caller asks for money, a Pin or a one-time Pin, hang up and call back on a number you already have saved.