There is a common definition of safety that applies to most of the world.

Which is that any AI is not racist, misogynistic, aggressive etc. It does not recommend to people that they act in an illegal, violent or self-harming way or commit those acts itself. It does not support or promote nazism, fascism etc. Similar to how companies deal treat ad/brand safety.

And you may think of it as a weasel word. But I assure you that companies and governments e.g. EU very much don't.