Multilingual AI Safety Benchmarking

ROK-FORTRESS is a new benchmark developed to evaluate AI safety across languages and geopolitical contexts, moving beyond translation-only evaluations. It systematically adapts linguistic form and contextual grounding to measure model behavior, revealing that language and context interact in complex ways and that translation-only benchmarks can misestimate real-world safety gaps. This contributes to the development of more realistic AI safety evaluations that reflect diverse deployment conditions.