What's this about?
Safety training for LLMs is mostly done in English, on the assumption it carries over to other languages. This paper shows that assumption is an illusion for low-resource languages — the guardrails don't reliably generalize, leaving exploitable safety gaps.
What exactly is "low-resource language" safety?
A low-resource language is one with relatively little training data online (many non-English, non-Chinese languages). Because safety alignment is trained mostly in high-resource English, a model may refuse a harmful request in English but comply when the same request is phrased in a low-resource language — a known jailbreak avenue.
Why it matters
As these models deploy globally, safety that only works in English isn't safety — it's a false sense of it, and a real vector for abuse in the languages labs test least.