Modern AI systems are deployed globally, across cultures and in hundreds of languages, yet most safety research and evaluation remains English-centric. In this talk, we will outline a pragmatic roadmap for scaling safety beyond a single linguistic or cultural frame. We will first outline AI safety as a full-stack technical discipline spanning robustness, alignment, privacy, misuse resistance, and critically, evaluation. We will then argue that harm is not universal: what counts as harmful varies with local norms and histories. Drawing on evidence from multilingual red-teaming and jailbreak studies, we will show higher failure rates in low-resource languages and the limits of translate-and-test approaches. We will introduce a global-vs-local harm lens, address data scarcity and long-tail challenges, and present actionable mitigations. Finally, we will examine fairness in model evaluation and close with concrete recommendations for building culturally aware benchmarks and auditing multilingual safety so models are not only capable, but reliably aligned with the communities they serve.