You don't. You handle it the same way you handle security holes which every LLM (including "safe" American ones) will put in your software sooner or later -- by understanding and reviewing the output.