What stops a small, or even a large group of people to intentionally "poison" the LLMs for everyone? Seems to me that they are very fragile, and that an attack like that could cost AI companies a lot. How are they defending themselves from such attacks?