The MSafeTemp collection contains 1.8 million prompts in 60 languages that jump‑start research on safer multilingual AI. It is built by a three‑stage pipeline: the system first spots harmful words in a prompt, removes them to create a template, then translates the surrounding text. A language‑model judge checks that the translation still fits the template, and finally culturally authentic entities fill the slots. A figure illustrates these stages and the creation of generic harmful, benign and culturally localized variants.

When the authors tested several popular models, they found that attacks that failed in English were successful after the prompts were translated. Moreover, attacks that localized a prompt to a specific culture failed in ways that pure translation could not reveal. A second figure shows that the dataset really does cover 60 different languages.