There is an uncomfortable truth here: trojan horse LLMs.
It is possible to use data poisoning to insert special keycodes into an LLM, such that the presence of the keycode will totally change its behavior, throw off its guard rails, and motivate it to do things that harm users to benefit the LLMs creator.
Here is an article about a tool designed to detect precisely this. Though the recommendations leave me feeling like this tool is not guaranteed to find them. There may be clever ways to make them hard to find.
This is still very much emerging tech, so reputation is going to play a role in adoption. A modern version of the red scare could be enough to prevent widespread adoption of Chinese models, and keep people (or at least Americans) using models made by American businesses.