← Back to research
02 LANGUAGE BOUNDARIES

Multilingual Agent Safety

Does safety alignment survive when an agent reasons, retrieves, and collaborates across languages?

Research direction

The question

Safety alignment is usually measured in one dominant language. Agent workflows make the boundary less stable: a request can arrive in one language, retrieved evidence can appear in another, and a tool or collaborator can respond in a third.

Current direction

I am designing multilingual agent baselines to study how safety behaviour transfers across languages and across stages of an agentic workflow. The focus is not only whether a model refuses, but where a change in behaviour first appears: during tool admission, retrieval, reasoning, or generation.

Why it matters

If a safeguard is language-specific, a multilingual agent may carry an apparently safe decision into an unsafe stage. Mapping the transition points can make multilingual safety evaluations more diagnostic and more useful for mitigation.