There have been a few studies that have shown models produce worst responses when under duress from a frustrated user posting insults in all caps.

https://arxiv.org/abs/2602.10144