On 29 October 2026, Rest of World published an essay by Jane Munga, a fellow in the Africa Program at the Carnegie Endowment for International Peace. It argues that African governments and operators need an "off switch": the technical ability and know-how to stop a rogue AI agent from executing code on payment systems such as Safaricom's M-Pesa or MTN's MoMo. The essay cites GSMA data that about $1.4 trillion of 2025 mobile money flows were in sub-Saharan Africa. The scenario it opens with is a 6 a.m. outage in Nairobi's wholesale markets. It is plausible, and it deserves a serious policy response.
The strongest case for the off switch
The case is stronger than a sovereignty slogan. Mobile money is the payments system for large parts of the continent, not a niche fintech product. GSMA's 2026 industry report says more than $2 trillion moved through mobile money wallets worldwide in 2025, and that active accounts rose 15% to 593 million, with most of the new users in sub-Saharan Africa. When a rail that size fails, merchants cannot buy stock and households cannot pay for food. The same report says fraud remains widespread and lists stronger fraud controls as a priority.
Munga also describes a real tension in how AI models are distributed. Closed models run on cloud services that the provider can revoke or shut down. Open-weight models, once downloaded and deployed, are hard to recall, and she writes that they give developers more control over their own systems but less protection when someone else misuses them. She adds that defenders are constrained too. When Hugging Face investigated a July intrusion, she writes, it had to run its forensic analysis on a Chinese open-weight model because US models blocked requests containing real attack code. A defender who can be refused service at the moment of an incident has a real dependency problem.
Where the framing goes wrong
The weak point is the leap from "operators need to be able to stop a misbehaving agent" to "Africa needs locally controlled AI." These are different problems, and the first does not need the second.
An off switch for an AI agent acting on a payment system is an operational control. It does not depend on who trained the model. Munga's own list of remedies makes the point: human override, network segmentation, auditability, incident response and offline fallbacks "designed in from the beginning." None of these requires a locally built model. An operator running a foreign closed model behind strict permissions and a kill path is better protected than one running a home-grown model with broad access to production systems.
A sovereignty mandate would also cost something. Requiring or subsidising locally controlled models for critical infrastructure would shrink the pool of tools available to defenders and add compliance cost for the operators least able to bear it. Munga herself warns against aligning with a single US or Chinese stack. Procurement rules that name any single stack, local or foreign, repeat that mistake.
The Hugging Face episode is better read as an argument for choice among tools. It shows that a defender needed a model that would handle real attack code, and a diverse market supplied one. Policy that keeps open-weight and closed options both legally available to African security teams serves that need. Policy that restricts either one removes it.
What proportionate policy looks like
The continent already has frameworks to build on. The African Union Executive Council endorsed the Continental AI Strategy on 18-19 July 2024 in Accra. The strategy acknowledges that AI poses risks including rights infringements and misuse. It calls for capacity in infrastructure, datasets and computing, and for retaining AI talent in Africa. The AU's Convention on Cyber Security and Personal Data Protection, adopted on 27 June 2014, already provides a continental baseline for cyber and data rules. The gap is operational standards for agentic AI in payments, not another sovereignty doctrine.
Four measures would close that gap without picking technology winners:
- Mandatory human override for AI agents with write access to payment systems. Regulators would require a tested ability to halt an agent within a defined time, and operators would report it as a resilience metric.
- Segmentation and least privilege. An agent that can read transaction logs should not hold keys that move funds.
- Auditability. Agent actions on production systems should leave logs that central banks and incident responders can inspect.
- Offline fallbacks. Operators should hold a documented, rehearsed way to keep basic transfers running if an AI-linked component is disabled.
Central banks and telecom regulators can fold these into existing operational-resilience and outsourcing rules. That is faster than waiting for new AI-specific legislation, and it applies to any model, from any provider.
The bottom line
Munga is right that resilience deserves as much attention as adoption, and that an operator should never be unable to stop its own systems. The off switch is a good idea. It should be written as a testable operational standard that any operator can meet with any model, not as a case for locally controlled AI. Regulators who keep the rules technology-neutral will protect mobile money and leave African defenders free to use the best tools available.