Don't Trust the Salt: AI Summarization, Multilingual Safety, and LLM Guardrails