1 paper
Mohammed Ahnouch, Lotfi Elaachack
A deployed language model may refuse a harmful request in English yet comply with its faithful translation, revealing a cross-lingual safety failure that cannot be characterized re…