The safety of OpenAI's GPT-4 gets lost in translation

By translating "unsafe" commands first into Zulu and other under-represented languages, Brown scholars coerced GPT-4 into breaking its guardrails.