AI agents increasingly rely on step-level guards to decide, before every tool call, whether the next action is safe given the user's request and the conversation so far. These guards are meant to catch harmful or unintended operations in real time. But the prompts that define guard behavior, and the benchmarks used to test them, are usually written in English.

The new paper, posted on arXiv, points out that users naturally write requests in many languages. When a guard's expectations are shaped by English-centric data, a request in another language may not trigger the same scrutiny. The authors frame this as a security surface: an attacker or even an ordinary user could bypass safety checks simply by switching languages.

The work does not claim that all guards fail on non-English input, but it raises a concrete concern for agent deployments. If safety decisions are language-dependent, then robust guardrails must be evaluated across the languages users actually speak, not just the language of the guard's own prompt. The paper stops short of proposing a full fix, but it makes the case that language coverage is a first-order security property for agentic systems.