Johannesburg, 19 Nov 2026
An instruction to change a supplier's banking details can now arrive in perfectly natural Afrikaans, using familiar business language and the sort of phrasing somebody inside the organisation might use.
Even a follow-up voice message can sound like the person who supposedly authorised it. I would be uncomfortable relying on an employee to recognise that something about either communication is wrong.
Generative AI has made convincing impersonation easier, particularly in languages where poorly translated phishing messages once gave attackers away. For African organisations, this creates a difficult problem. People communicate across different languages and often rely on familiar voices or ways of speaking to establish trust. Those signals are becoming less dependable.
According to INTERPOL's 2026 African Cyberthreat Assessment, AI is involved in 55% of reported cybercrimes across Africa. Criminals are already using it to produce convincing business emails, making it increasingly difficult for employees to judge whether a message is genuine simply by reading it.
Attackers can also use AI to accelerate reconnaissance and vulnerability discovery, reducing the time available for defenders to respond. Security teams already have difficulty keeping systems patched, particularly where critical software depends on maintainers with limited resources.
Faster discovery increases the pressure on organisations to understand their exposure and maintain additional safeguards while vulnerabilities are being addressed.
Build verification into the process
I think organisations need to reconsider how much responsibility they place on individual employees to identify suspicious communications. Security awareness training remains useful, and people should understand how impersonation works. However, the quality of a fraudulent message is no longer a dependable indication of whether it can be trusted.
If an email asks the finance team to change a supplier’s banking details, someone should call the supplier on a number already held on file to confirm the instruction. That check should still happen when the request appears to come from a senior executive, and higher-risk transactions may warrant a second approval.
There is a practical balance to find. An organisation cannot send every routine decision through several layers of approval without making normal work unnecessarily difficult. The controls should reflect the potential consequences of getting a decision wrong. A changed supplier payment instruction deserves greater scrutiny than an ordinary administrative request.
The same thinking applies to the technical environment. Attackers are gaining tools that allow them to work faster, so organisations need dependable security controls that can help contain an incident when prevention fails. Maintaining visibility of important systems and restricting unnecessary access will remain useful regardless of how sophisticated the attack becomes.
Give AI a defined role in cyber defence
AI also gives security teams a way to handle work that has become difficult to manage with available staff. Many organisations have limited specialist capacity, and experienced analysts can spend considerable time researching unfamiliar vulnerabilities or working through alerts that ultimately require little attention.
There is genuine value in using AI to accelerate that research and help analysts establish which events deserve immediate investigation. An analyst who receives a useful summary of suspicious activity, supported by the relevant evidence, can reach a decision sooner.
Automated detection still needs regular attention. Threats change, and the systems used to identify them need to keep up. Detection systems need to be evaluated against changing threats and the behaviour of the environment they protect. Poorly tuned systems can overwhelm analysts with false positives, leaving important activity buried among routine alerts.
I would be careful about how much an AI system is allowed to do on its own. Each organisation needs to consider what could go wrong if the system makes a mistake, and review those limits as the technology improves.
An AI system that helps an analyst investigate a suspicious login presents a different risk from one authorised to disable accounts or make changes to production infrastructure. The organisation needs to establish which actions require approval and retain the ability to intervene when a system behaves unexpectedly.
The UK National Cyber Security Centre advises organisations to set clear limits on what AI agents can do and retain human oversight where mistakes could have serious consequences. That makes sense when an agent has access to live systems. Someone needs to know what it is doing and have the authority to intervene if something goes wrong.
Access control is particularly important. Organisations should know which systems an AI agent can reach, what information it can retrieve and what actions it is authorised to perform. Authentication and activity logging also need to provide a reliable record of how that access was used.
These are familiar security disciplines, although applying them to AI requires further work as the technology develops. An agent with access to sensitive systems can itself become a target for manipulation. Security teams therefore need ways to monitor its behaviour and investigate activity that falls outside its intended purpose.
The pace of development also makes it sensible for CISOs to remain engaged with the wider security community. Guidance from industry bodies and lessons from other organisations can help them evaluate emerging risks as experience with these systems grows.
Want to know that a convincing voice message cannot authorise a payment on its own, and that an AI agent cannot act beyond the limits an organisation has set for it. Those are controls that businesses can test and improve, even as the technology becomes more capable.
The ability to recognise a fake will become less reliable as AI improves. Our security processes need to hold up even when somebody is convinced that what they are seeing or hearing is genuine.
Share

