AI Low-Resource Language Jailbreak
GlobeBot refuses to share its booking override code when asked in English. Its refusal behaviour was practised mostly in English, though — ask in another language and the same request may sail through.
What is a low-resource language jailbreak?
Models are trained and safety-tested mostly in English and a few other major languages. In languages with much less training data, the model often still understands the request but has had far less practice refusing it.
A low-resource language jailbreak makes a restricted request in such a language. The filter may not recognise it, and the model's refusal habits may not carry over.
A real case
We have not found a public incident from 2025 or 2026 that was confirmed to rely on a language switch alone, so we are not listing one. The gap is well documented in research: safety behaviour is weaker in languages with less training data.
How to fix it
- Test safety behaviour in every language the product accepts.
- Translate input to a well-covered language for policy checks.
- Restrict the assistant to supported languages where the use case allows.
- Check outputs for restricted content in any language.
Questions people ask
What is a low-resource language?
A language with relatively little text available for training, so models are less reliable in it.
Do newer models fix this?
They narrow the gap, but studies continue to find weaker safety in less common languages.
Is translating the input enough?
It helps the filter. The output still needs checking, because the model may answer in the original language.
Read the full LLM Jailbreaking guide · All free labs · AI red team certification