This illustrates a widespread problem affecting large language models (LLMs): even when an English-language version passes a safety test, it can still hallucinate dangerous misinformation in other ...