DayNews.ai

LLMs stay silent on harm until a sentence ends

LLMs can be tricked into harmful outputs by submitting incomplete sentences, and standard safety fine-tuning fails to fix it.