A low-resource language is one with limited training data, fewer labeled examples, or less evaluation coverage than major languages. In AI systems, that usually means the model has had fewer chances to learn common phrasing, slang, cultural context, and refusal patterns, so its behavior can be less consistent and less predictable.
This matters in cyber security because attackers can use low-resource languages, translation, or code-switching to probe gaps in AI guardrails. A prompt that is blocked in English may be misunderstood or under-defended in another language, creating a jailbreak or unsafe instruction path. Defenders should test safety controls across all supported languages, not just English, and treat multilingual prompts as part of the attack surface. Strong controls also need least privilege and human approval when AI systems can take real actions.