AI safety guardrails, the mechanisms designed to block jailbreaking attempts and prevent unsafe outputs, do not perform consistently across languages, according to new reporting on AI security testing. The finding carries particular weight in Europe, where organizations routinely operate across dozens of languages and dialects, creating a wider attack surface than English-only deployments.

Most large language models and the security layers wrapped around them are developed and tested predominantly in English. As a result, the guardrails meant to catch malicious prompts, block harmful content generation, and prevent unauthorized actions can be far less reliable when the same techniques are attempted in other languages. A jailbreak prompt that gets flagged and blocked in English may slip through undetected when phrased in French, German, Polish, or dozens of other languages spoken across the continent.

Why This Matters for Security Teams

For enterprises deploying AI chatbots, copilots, and autonomous agents across multilingual user bases, this inconsistency represents a genuine security gap rather than a theoretical concern. Attackers do not need sophisticated tooling to exploit the weakness. Simply switching the language of a malicious prompt may be enough to bypass protections that would otherwise catch the same request in English.

This is especially relevant for organizations in the European Union, where regulatory and business realities often require AI systems to function across many official languages. A security posture that only accounts for English-language threats leaves those deployments exposed.

What Organizations Should Consider

  • Test AI guardrails and content filters explicitly in every language your deployment supports, not just English.
  • Treat multilingual jailbreak attempts as a distinct threat category during red-teaming exercises.
  • Push AI vendors for transparency on how their safety layers are trained and validated across languages.
  • Monitor for anomalous non-English prompts that may indicate probing or jailbreak attempts.

As AI adoption accelerates across multilingual regions, security teams should not assume that guardrails proven effective in one language automatically extend to others. Closing this gap will likely require vendors to invest more deliberately in multilingual safety testing, and for enterprises to build that expectation into their AI risk assessments.