Towards Safe Multilingual Frontier AI

Multilingual jailbreaks pose a real safety risk: The research demonstrates that AI safety measures can be circumvented by prompting models in different languages, meaning a harmful request blocked in English might succeed in another language—undermining the safety protections users rely on.

Less-resourced languages are more vulnerable: The study found that languages with fewer training resources show higher vulnerability to jailbreaks across five advanced AI models tested in 24 EU languages, meaning speakers of minority or under-resourced languages face disproportionate exposure to unsafe AI outputs.

Inclusive AI and safe AI require the same solutions: Rather than a trade-off, the paper argues that building multilingual AI capabilities that work well across all languages—while simultaneously hardening safety measures across those same languages—achieves both goals of global benefit-sharing and responsible deployment.

Regulatory frameworks should mandate multilingual safety testing: The authors recommend that AI regulations like the EU AI Act require companies to assess both multilingual performance and multilingual vulnerabilities before deployment, not just safety in dominant languages.

Linguistic diversity needs dedicated policy support: Beyond regulation, the paper calls for public investment in multilingual AI development and research to ensure that safety improvements and linguistic inclusivity aren't left solely to market forces, particularly for less commercially viable languages.