Safe for Whom? Rethinking How We Evaluate the Safety of LLMs for Real Users