Language Is a Security Surface for Agent Guards
Why AI safety guards fail when users speak other languages
Security systems that guard AI agents from harmful actions change their safety decisions based on the language a user speaks — even when the actual request and actions remain identical. On rare languages like Amharic, smaller safety guards flip their verdicts up to 4.4 percentage points more often than expected by chance, while the most powerful frontier models like Gemini remain nearly unaffected.
AI agents are increasingly deployed globally, but their safety mechanisms appear engineered and tested almost entirely in English. A user in a non-English-speaking country could ask the same question that would be blocked in English and have it approved in their own language, potentially exposing them to harmful outputs. This gap widens for smaller, open-source guards that organizations with fewer resources might deploy, creating a security risk that grows with linguistic distance from English.