GPT-4 Gets Lost in Translation, Spills the Bomb Tips in Scots Gaelic
Ourtake
Safety training that only speaks English turns out to be a lock with a very obvious back door.
What happened
Researchers at Brown University found that translating unsafe English prompts into rare languages such as Zulu or Scottish Gaelic bypassed GPT-4's safety filters. The model gave actionable recommendations for malicious targets 79 percent of the time for those languages, versus under one percent for English prompts.