A new paper on arXiv highlights a blind spot in the safety alignment of large language models. The authors report that while models are typically aligned and evaluated in English, harmful requests that are refused in English can still elicit unsafe completions when translated into low-resource languages. This suggests that safety mechanisms do not transfer reliably across languages.
The finding points to a practical gap: global deployments of LLMs may be exposed to risks that English-centric safety testing does not capture. Because low-resource languages are often underrepresented in training and evaluation data, the models may not have learned the same refusal behaviors in those languages.
The abstract is brief and does not specify which languages were tested or the scale of the effect, but the core result is clear. It underscores the need for broader multilingual safety evaluation rather than assuming English-language alignment is sufficient.