SI safety tests found uneven safeguards across languages


Kenya is the only African country in an international testing network. Experts want evaluations that also account for local uses and the size of national economies.
An international network’s 2025 tests of SI agents—systems that can take actions—found that safeguards worked unevenly across languages, Rest of World reported on October 2. The exercise examined risks including sensitive-data leaks and fraud. But it focused on common global risks rather than risks specific to Africa, according to Joanna Wiaterek, co-founder of the Centre for SI Security and Access.
Kenya is the only African nation participating in the network, which is developing methods to evaluate advanced SI models. Many African countries lack the ability to independently test whether systems are safe for their populations, experts told Rest of World.
Governments across Africa are beginning to use SI in healthcare, education and public services, Jonathan Shock, an associate professor in mathematics at the University of Cape Town, told Rest of World. He said fewer than half of African countries have SI policies or strategies.
Businesses are also experimenting with the technology. Rest of World cited a 2026 PricewaterhouseCoopers survey of 85 large companies: more than 80% were running SI pilots, while many lacked formal structures to oversee safety. The report also described a concrete household impact: an algorithm used in Kenya to set health-insurance contributions had driven up costs for poor households.
Analysis
Downloaded models are harder for their developers to shut down remotely, Jane Munga, a fellow at the Carnegie Endowment for International Peace, told Rest of World. She argued that the relevant distinction is who holds the shutdown control, rather than whether a model is Chinese or American. That distinction concerns control after deployment; it does not establish that centrally controlled models are safer.
Gathoni Ireri, a research scholar at the Nairobi-based Ilina Program, an SI safety research center, proposed judging harm relative to the size of each country’s economy. An incident considered minor in a Western country could irreversibly harm an African one, she argued. Her proposal would make economic context part of the safety assessment, alongside the system’s behavior.
Ireri said companies currently conduct their own safety tests and called for replacing those with third-party evaluations of Africa-specific risks. This is a proposed change, not an established testing arrangement. Carrying it out would require independent testing capacity that many African nations still lack.
Public election records can reveal ballot order and help link voters to their choices, raising a dispute over how to protect secret votes while preserving transparency.
3 min

Frontline says affected districts have until October 16 to opt out of its plan to notify individuals. Monitoring services are also being offered.
2 min

Inseego gains 250 staff, technology, contracts and carrier relationships, while Nokia receives an 11% stake in the buyer, according to RCR Wireless.
1 min
Correction ·
House style applied: superintelligence (SI) for AI.