Large language models are deployed worldwide, yet safety alignment and evaluation remain focused on English. The study shows that harmful requests rejected in English can produce unsafe completions when translated into low‑resource languages. Under a translation‑bypass threat model we evaluated ten widely used models across 79 languages (English plus 78 low‑resource languages) using a metric conditioned on English safety, which separates translation bypass from unsafety already present in English.
The findings indicate that English safety does not transfer reliably: when English completions are safe, the average bypass rate ranges from 21.9% to 44.4%. Bypass risk clusters in certain language families, a pattern shared across all models. Although this ordering is confounded with translation degeneration, the variation between models is significant and persists under every control we applied.
Further analysis reveals that per‑model bypass rates do not yield a stable ranking: nine of ten models change rank across three conditioning schemes that estimate the same quantity, even though the underlying responses remain unchanged.
Review: Translation bypass exposes a critical language bias in safety evaluation, highlighting the need for multilingual safety mechanisms to protect users of low‑resource languages.