HausaSafe

AI safety evaluations, red-teaming, benchmarks, and alignment testing are built almost exclusively in English. This isn't a minor gap; a model can pass every English-language safety benchmark while behaving unpredictably in Hausa, spoken by 60+ million people across northern Nigeria, Niger, and the Sahel, where AI systems are already deployed in health, finance, and security infrastructure with zero language-specific safety evaluation.

HausaSafe gives AI labs and safety researchers the tools to evaluate model behaviour in Hausa for the first time. This is grounded in the academic argument of Ireri, Abungu et al. (2026) that Africa is more likely to expose universal AI failure modes through distributional shift than to generate entirely distinct pathways of misalignment. This means Hausa evaluation doesn't just protect Hausa speakers; it surfaces failure modes the global safety community cannot currently find in English at all.

Ungoverned AI deployment at scale in low-resource, non-English-language contexts is an attack surface that the current safety architecture cannot see. Closing that blind spot, even partially, reduces the risk that catastrophic failure modes go undetected simply because nobody built the tools to look for them outside English.