Untapped AI, a company based in Cape Town, filed two patent applications in the US in September. These applications relate to technology capable of reliably determining the source of a phone call. The goal of the development is to counter criminals who use artificial intelligence to clone voices and impersonate bank representatives, executives, or family members.
Although the technology has been tested, it is not yet operational in real-world conditions. Protection will only be available to users after it is implemented by mobile networks or commercial entities on both ends of the call.
Untapped AI creates AI-based voice agents for businesses and has offices in Johannesburg and Cape Town. The company received investment from telecommunications and contact center provider Yaxxa, whose CEO Heinz Vollmer confirmed the investment.
Technology Principles
Director Lloyd Matthew stated that the technology consists of two components, which he did not disclose in detail for security reasons. Both elements include a cryptographic key that changes with every call. This key is either exchanged at the network switch upon connection or transmitted within the audio signal as a signal imperceptible to humans.
The second part of the system analyzes sound frequency, pauses, and other signal markers to verify the integrity of the conversation, rather than simply comparing the caller's voice to a recorded sample of the recipient's voice. According to Matthew, the system determines the origin of the voice, meaning the source of the call.
He compared the operating principle to a bank card: attempting to bypass protection using deepfake is similar to a person trying to withdraw cash from an ATM using a stolen card while wearing the owner's mask. The machine checks the PIN code, not the face.
Financial Losses and International Experience
The company's press release describes the 'voice authentication in a phone call' technology, but in written responses to TechCentral, Matthew asserted that the patents 'do not analyze the voice at all.' When asked about the discrepancy, he noted that 'both statements are true, and the wording could have been clearer.'
According to South Africa's Banking Risk Information Centre (Sabric), South Africans incurred losses of 2.4 billion rand in 2025 due to digital banking crimes in over 110,000 registered cases. Sabric specified that it cannot determine what proportion of these losses is specifically related to phone calls. It acknowledged that AI voice cloning is an emerging risk, but 'current data does not indicate that it has displaced traditional social engineering schemes in South Africa.'
Abroad, the use of cloned voices is already causing significant losses. In February, fraudsters withdrew 95 million euros from the private Italian bank Fideuram, which is part of Intesa Sanpaolo. They used a fake WhatsApp message followed by a phone call where AI was used to imitate the voice of a senior lawyer, according to Reuters. More than half of this amount has already been recovered.
When asked whether the technology protects anyone before its installation, Matthew replied negatively: 'No. The protection must be verified on both sides of the connection, so it starts working as soon as the network or business deploys it on each end. That is why our initial negotiations are with networks, as one such deployment covers all subscribers.'
Untapped AI has entered into negotiations with a major mobile network, which he did not name. Matthew emphasized that the company deliberately does not disclose the process in detail or publish test results because 'we are intentionally vague on this matter because we do not want to publicly describe the actual process.'
Professor Martin Olivier from the University of Pretoria's Faculty of Computer Science stated that he could not fully assess the claim based on the available information. He noted the well-known principle of cryptography: 'secrecy often indicates something in the system that needs to be hidden.' He added that all details of trusted cryptographic protocols, such as AES and RSA, are publicly available. He believes that a properly used, verified cipher will stop cloning, but warned that networks compressing audio 'may distort signals.'
In the US, networks use a call authentication standard called Stir/Shaken, which signs the caller ID number, not the audio signal itself. Olivier compared this to certificates that ensure trust in websites. Matthew countered that his system does not depend on the number displayed on the screen.
Andris Marits, Product Manager at Entersekt, noted that embedding a signal within the audio signal 'is a significant step forward under certain conditions.'
Non-Autonomous Solution
Marits explained that the signal embedded in the audio is tied to the call itself and cannot be easily separated. However, the network can only check calls that pass through it. A fraudster using VoIP or another route can bypass this check, so banks still require additional controls. VoIP refers to calls made over the internet.
Sabric believes that call authentication can help close a gap in telephone communications, but stressed that it should not be seen as a standalone solution. He warned that criminals adapt quickly and 'may increasingly use alternative channels such as WhatsApp, SMS, email, or social media platforms.'
Chief Executive Officer of Banking Technologies George Chinomona, speaking in a personal capacity, stated that AI mainly improves the operation of old scam schemes. He advised banks to assume that voice alone can no longer serve as proof of identity.
Matthew expresses hope that the government will change the distribution of responsibility for damage caused by fraudulent calls. 'We hope the government will develop a policy whereby responsibility for call fraud shifts from the client to the business. If this happens, we will share our code with the market, including competitors.'
MTN stated that any national structure must be created jointly by the industry, regulators, and technology providers. The company noted that open and compatible standards 'promote wider adoption, innovation, and uniformity across networks.'
Until any of these systems are launched, Matthew's advice remains simple: agree on a family code word. If the caller demands money, a PIN, or a one-time PIN, hang up and call back using a number you already have saved.
For Marits, the hardest task is not the technology itself. He believes: 'The technology is less important than ensuring that enough of the ecosystem adopts it simultaneously.'
