LLM agents increasingly use retrieval tools across languages, yet their safety evaluation remains English-centric. A new paper on arXiv (2610.23093) asks whether language affects tool admission, post-retrieval generation, or both, rather than assuming safety performance transfers across languages.
The authors propose to independently examine these two stages. Tool admission refers to whether the agent decides to invoke a retrieval tool, while post-retrieval generation concerns how the agent responds after receiving tool output. This distinction matters because safety failures could originate at either point, and language may influence them differently.
The abstract does not include results, so the paper's conclusions are not yet available from the source. Still, the work highlights a potential blind spot: if safety is only tested in English, language-specific failures in tool use or generation could go unnoticed.