And this model is arguably their least impressive model.

Mostly by name association. The LLMs named Grok are good LLMs. The twitter bot of the same name, using those models and a custom prompt, has a habit of creating controversy. Usually after somebody modified the system prompt.

I use grok a lot on the web interface (grok.com) and never had any weird incidents. It's a run-of-the-mill SOTA model with good web search and less safety training

How does somebody modify the system prompt over an x message to the chat bot?

I think somebody here refers to a very specific meddling CEO