Some AI safety guards judge an identical tool call differently when the user's request changes language. Kontext researchers Hasan Demirkiran and Jens Ernstberger report the finding in a study published on October 15.
LinguaGuardBench translates requests into German and four other languages while retaining 1,220 stored action steps. Some smaller guards missed more unsafe steps; others blocked more harmless ones.
The tested frontier models remained largely stable. The authors argue that English-only evaluations can miss these differences.
No tools were executed. The results concern these test configurations, not demonstrated real-world attacks. Two languages lacked native-speaker review, and the statistical uncertainty may be larger than reported.