When Seeing Is No Longer Believing: Can Zambia’s Treasuries Survive the Deepfake Payment Order?

In January 2024, a finance employee at the Hong Kong office of Arup, the British engineering firm, joined a video call with the company’s chief financial officer and several colleagues. He was asked to make confidential transfers, and he did: fifteen of them, worth about US$25 million. Every other face on that call was a deepfake. Nothing had been hacked and the firm’s systems worked exactly as designed. What failed was the oldest control in finance, i.e., recognising the person giving the instruction.

That incident should unsettle every treasury in Zambia. Across the world, large payments are still released on the strength of a familiar voice, a video call or an email from a senior executive. Generative AI has made each of these cheap to forge. The question is no longer whether such attacks will reach Lusaka, but whether our controls will notice when they do.

A control built for a world that no longer exists

Payment authorisation rests on a simple assumption, which is, that seeing and hearing someone is proof of who they are. Research shows how fragile that assumption has become. A major survey of face manipulation found synthetic faces and expressions increasingly difficult for people to distinguish from genuine recordings. Criminals have noticed. The United States Financial Crimes Enforcement Network reported a rise in suspicious activity reports describing deepfake media used to circumvent identity verification, and warned that the same tools can impersonate trusted individuals in scams against businesses.

The losses are already large. The FBI’s 2025 Internet Crime Report recorded more than US$20.8 billion in reported cyber-enabled losses, including over 22,000 complaints referencing artificial intelligence that accounted for nearly US$900 million. Business email compromise, the attack class that deepfakes now supercharge, remains one of the costliest categories year-after-year.

Why better detection will not be enough

The instinctive response is to buy a deepfake detector. Detection research is advancing, but it is brittle. The benchmark FaceForensics++ study showed that classifier accuracy falls sharply on compressed video, precisely the kind carried by congested, low-bandwidth networks. Every improvement in detection also becomes a target for the next generation of forgeries. A control that depends on spotting the fake will always be one release behind.

Africa: industrialised fraud, thin defences

The risk is rising fastest where defences are thinnest. INTERPOL’s African Cyberthreat Assessment Report 2026 found that artificial intelligence enabled 55 per cent of reported cybercrimes on the continent in 2025, while cybercrime legislation remains fragmented and AI readiness among law enforcement agencies is alarmingly low. Executive impersonation is not hypothetical here. During Operation Sentinel in late 2025, a Senegalese petroleum company detected fraudsters who had infiltrated its email systems and impersonated executives to authorise a wire transfer of US$7.9 million.

Zambia: the sharpest rise on the continent

Zambia sits at the centre of this shift. Sumsub’s Identity Fraud Report 2025–2026 recorded a 967 per cent surge in deepfake attempts against identity verification in Zambia, the highest growth in Africa, and found that the country stood out for the number of applicants involved in fraud networks. The consequences reach the balance sheet. A University of Zambia study found that fraud losses significantly reduce bank return on assets.

The legal framework has moved, but not far enough. The Cyber Security Act, 2025 and the Cyber Crimes Act, 2025 create incident-reporting duties and offences for cyber-enabled deception, yet neither prescribes how a treasury should verify an instruction before money leaves the account. That decision is left to each institution, and many still rely on a call-back to a number the attacker may already control.

Can an AI check the AI?

One promising answer is to make verification independent of human perception. Agentic AI systems, which reason and act through external tools in iterative loops, are maturing quickly, as a recent survey of language-model agents documents. An agent could sit between instruction and release, checking each payment against invoice history, counterparty records, behavioural baselines and an out-of-band challenge to a pre-registered device. It would not ask whether the face looks real; it would ask whether the request makes sense.

But agents bring their own attack surface. Researchers have shown that language-model applications can be hijacked through indirect prompt injection, where malicious text hidden in an email or document quietly rewrites the agent’s instructions. A verifier that can be talked into approving a payment is worse than none at all.

What it would take

Four conditions should govern any deployment. First, segregated authority, i.e., no single person, and no single agent, should be able both to request and to release a payment. While banks and the more sophisticated corporate treasuries and finance teams generally have dual controls or ‘maker-checker’ processes embedded in their procedures, a useful step would be having this as a mandatory requirement at the point of payment instruction origination. Second, tamper-evident logs, so that every decision can be audited after the fact. Third, bounded tool access, so the agent can read records and raise challenges but never move money itself. Fourth, human escalation for anything unusual, against clear thresholds. A practical first step would be a controlled pilot with two or three Zambian banks and corporate treasuries, simulating deepfake attacks against real workflows and measuring what is caught, what is missed and at what cost.

The question is not whether forgeries will become perfect. It is whether Zambia’s payment controls can remain trustworthy once seeing and hearing are no longer believing.

___________________

**RIS Contributors are researchers and practitioners writing for CCET’s Research Intelligence Service, views are the author’s own.

Sources

Primary sources

•         Federal Bureau of Investigation (2026). 2025 Internet Crime Report. Washington, DC: FBI Internet Crime Complaint Center

•         Financial Crimes Enforcement Network (2024). FinCEN alert on fraud schemes involving deepfake media targeting financial institutions. FIN-2024-Alert004, 13 November 2024. Vienna, VA: FinCEN

•         Government of the Republic of Zambia (2025). The Cyber Security Act, No. 3 of 2025. Lusaka: Government Printer

•         INTERPOL (2025). 574 arrests and USD 3 million recovered in coordinated cybercrime operation across Africa. News release on Operation Sentinel. Lyon: INTERPOL

•         INTERPOL (2026). INTERPOL report finds AI linked to more than half of cybercrime in Africa. News release on the African Cyberthreat Assessment Report 2026, 3 August 2026. Lyon: INTERPOL

Media and industry reporting

•         Bowmans (2025). Zambia: New laws to strengthen cybersecurity and cybercrime capacities. Legal briefing on the Cyber Security Act and Cyber Crimes Act, 2025

•         BusinessDay (2025). Africa enters high-skill fraud era as deepfakes surge 967% in Zambia, 367% in DRC. Reporting the Sumsub Identity Fraud Report 2025–2026, 2 December 2025

•         Infosecurity Magazine (2025). AI and deepfake-powered fraud skyrockets amid identity fraud stagnation. Reporting the Sumsub Identity Fraud Report 2025–2026

•         Magramo, K. (2024). Arup revealed as victim of $25 million deepfake scam involving Hong Kong employee. CNN, 16 May 2024

Research

•         Greshake, K. et al. (2023). Not what you’ve signed up for: Compromising real-world LLM-integrated applications with indirect prompt injection. Proceedings of the 16th ACM Workshop on Artificial Intelligence and Security, pp. 79–90

•         Rössler, A. et al. (2019). FaceForensics++: Learning to detect manipulated facial images. Proceedings of the IEEE/CVF International Conference on Computer Vision, pp. 1–11

•         Siasulingana, M. and Haabazoka, L. (2024). A study of the effects of mobile banking services on the financial performance of Zambian commercial banks: A case study of Atlas Mara. Social Science Journal for Advanced Research, 4(2), pp. 22–34

•         Tolosana, R. et al. (2020). Deepfakes and beyond: A survey of face manipulation and fake detection. Information Fusion, 64, pp. 131–148

•         Wang, L. et al. (2024). A survey on large language model based autonomous agents. Frontiers of Computer Science, 18(6), 186345

•         Yao, S. et al. (2023). ReAct: Synergizing reasoning and acting in language models. Proceedings of the Eleventh International Conference on Learning Representations (ICLR 2023)