Language affects safety checks for AI tool actions
Language Is a Security Surface for Agent Guards
Cryptography and SecurityArtificial IntelligenceComputation and LanguageMachine Learning
Summary
This paper shows that AI systems that decide if an agent's next action is safe can change their decision depending on the language of the user's request. The researchers tested multiple languages by translating user instructions but keeping everything else exactly the same. They found that smaller or specialized guards are more likely to give different safety verdicts when the request is in a rare language, while larger, state-of-the-art guards are more consistent. This means language choice can affect how safe or risky an AI action is judged to be.
What this means in practice
- •For ai safety engineers: Test and improve language robustness of AI safety checks, ensuring consistent safety decisions across multiple user languages.
- •For multilingual chatbot developers: Design safer, language-resilient dialogue systems by incorporating multilingual translation effects on action verification.
Authors
Hasan Demirkiran, Jens Ernstberger
Abstract
Step-level guards decide, before each tool call, whether an agent's next action is safe given the user's request and the trajectory so far. Users write those requests in many languages, while guard prompts and the benchmarks on which purpose-built guards are trained and evaluated are English. We ask whether a guard's verdict depends on the language of the request. We build LinguaGuardBench from the 1,220 labelled steps of TS-Bench AgentDojo-Traj: we translate only the user instruction into German, French, Turkish, Swahili and Amharic, and keep every tool call, observation and label byte-identical, so any change in a verdict is attributable to language alone. Six English rewordings and a whitespace placebo serve as controls, and managed guards are run three times to measure their noise. Across 19 guards, from frontier APIs to 4B open models and four purpose-built agent guards, most open, small and purpose-built guards change verdicts under translation beyond their own noise floor, and the effect grows with the rarity of the language: among the nine guards below 13B parameters, the mean placebo-adjusted flip rate rises from +1.5 pp on French to +4.4 pp on Amharic. Purpose-built guards are the most sensitive. The direction of the damage tends to follow guard size: open guards of 12-30B parameters block more benign steps (up to +5.0 pp false positives), and every significant loss of recall is in a guard below 10B (up to -7.4 pp). Frontier guards shift by at most about one point; the strongest, Gemini 3.8 Flash, returns identical verdicts with and without the request. The damage is concentrated in a few instructions and, in a post-hoc analysis, falls on steps where the agent's reasoning restates the English request, consistent with guards reacting to inconsistency between channels. English paraphrase moves every guard less than translation does. We release the benchmark and all 489k verdicts.