It’s not maybe they have been researched and are highly manipulable.

You are trusting the host of the model to not introduce biases. Getting the model to regurgitate it’s hard coded prompt info has already happened and different providers have been doing that