Is there one LLM 'command language' or many?

A given LLM tends to be more susceptible to certain kinds of prompts than others. For instance, a classic early jailbreak was having the model write a story that included the bad behavior being requested. The big question to me is: do these jailbreaks tend to be 'all encompassing', so that any bad behavior the model is ever likely to do can be accessed by a single jailbreak? Or is it that different behaviors are more susceptible to different jailbreaks? The really big question here is: How should we categorize behaviors to get a clean answer?