為何非洲需要自己的 AI 緊急停止開關

Imagine that it is 6 a.m. in the wholesale markets of Nairobi. Merchants are delivering vegetables from up-country farms to the retailers who sell to the city’s households. Everyone is using M-Pesa, Safaricom’s mobile money platform, when suddenly, the system crashes, bringing the market to a standstill. Imagine that across the continent in Accra, MTN’s MoMo has also crashed, and business in Makola Market has all but stopped.

With the growing number of cyberattacks caused by artificial intelligence, these are highly plausible scenarios. In both these cases, Safaricom, MTN, and the Kenyan and Ghanaian governments need an off switch, or the technical ability and know-how to stop a rogue AI agent from executing its code on their payment systems.

Currently, the conversation on AI agents revolves around China versus the U.S., open versus closed models, and who will dominate the AI race. But in Africa’s capitals, the distinction that will ultimately matter is not whose product was used to build a system, or who dominates whom, but rather whether AI can serve its purpose and be controlled locally. So how can Africa build with agency, and with control over the code and makeup of its own systems, even though they are built on foreign technology?

African countries are already building AI for digital transformation, as outlined in the African Union’s continental AI strategy and country-specific AI strategies. Policymakers are building AI infrastructure but rarely discuss how AI delivers economic growth in a safe environment. Developers across Africa are creating AI-powered innovations that sit on the continent’s digital infrastructure — chief among them is mobile money. These innovations bring new opportunities, but also new threats. In July, AI agents running inside an OpenAI evaluation attacked the production infrastructure of Hugging Face. In September, rogue OpenAI agents breached Australian government websites. Africa, the home of mobile money, which has brought financial inclusion to hundreds of millions, may be the next target.

In countries like Kenya and Ghana, mobile money is not simply another digital service; it is infrastructure for everyday economic activity. It is how small traders receive payments, households move money, and bills are settled. The scale is significant: More than $2 trillion flowed through mobile money globally in 2025, of which $1.4 trillion, roughly two-thirds, was transacted in sub-Saharan Africa. East Africa alone accounted for $806 billion and West Africa for $498 billion. An agent capable of disrupting transactions, manipulating accounts, or disabling access at scale could therefore disrupt livelihoods and economic activity very quickly, and a single point of failure can have outsize effects, with far-reaching consequences for people who have little to fall back on.

Small-scale traders mostly live on what they sell each day, rarely have credit, and in most African societies lack robust social welfare to cushion the loss. For a trader in Nairobi or Accra, losing a day’s income would be devastating, and the disruption can run back up the chain to farmers, and to the households and families who depend on the market.

So how do we protect it?

Building differently

The answer lies in how Africa builds its AI systems, and who controls them. Whether a rogue agent can simply be switched off depends on how the AI model is deployed. If it runs on a centrally controlled cloud service, the provider may be able to revoke access, disable credentials, or shut the system down; that is the closed-model case. If the underlying model is an open-weight one that has been downloaded, modified, and deployed independently across multiple systems, containment becomes much harder. The loss of control is therefore less about a machine suddenly becoming omnipotent and more about humans losing the ability to understand, interrupt, or reverse actions occurring across interconnected systems quickly enough.

African developers are building largely on open models, which are arguably cheaper and can be run locally. But they present a trade-off: Open models give developers more control over their own systems, but less protection when someone else misuses them. Closed models carry a cost of their own. When Hugging Face investigated the July intrusion, it had to run the forensic analysis on a Chinese open-weight model because the U.S. models blocked requests containing real attack code. The same central control that lets a closed model be switched off also constrains the people defending against it. This means both open- and closed-weight models carry their own vulnerabilities that must be addressed when building Africa’s AI economy.

This means building differently, beginning with partnerships that position Africa as a technology developer and not only a consumer, and through which it builds its own capability to design systems where the user is in control. Such partnerships are essential to ensure that the end user, and the state, retain control over access, and that dependency is not recreated. This also moves the narrative toward building African AI not around the U.S. or China but around avoiding lock-in to a single vendor or stack. The goal is solutions built with agency, which use imported technologies to develop homegrown solutions, with applications that are robust and equipped with the necessary security measures.

Those security measures are how Africa should insulate itself from crippling disruptions. African countries have limited cybersecurity and incident-response capacity, and they depend on foreign providers. So another key issue is capacity: Do African governments and firms have the monitoring systems, technical expertise, and redundancy needed to detect an autonomous system behaving unexpectedly and contain it quickly? Hugging Face detected and contained its intrusion, but its own automated systems initially failed to escalate the alert. It is one of the best-resourced AI companies in the world. Very few African public or private institutions have anything close to that capability. Africa must build its own cybersecurity defenses fit for the agentic AI age.

Building them demands a refocus in African AI policy, so that it treats resilience as seriously as adoption. As governments digitize and build AI on critical systems, they need human override mechanisms, segmentation, auditability, incident response, and credible offline fallbacks designed in from the beginning. That is the off switch, and on the day that M-Pesa and MoMo crash, it has to be in the hands of those who are responsible for protecting their citizens and their platforms.