With local LLMs you can edit the LLMs response to e.g. something really vile and then ask why it would say this. Depending on the model (e.g. Gemma) it apologizes a lot but also promises that it’s a bug and people are working on it.

Llama in a psychology role play sometimes doubles down on it and says it said it to get an emotional response and even keeps using some of the vile stuff.