More like: the training data for LLMs is full of people moralizing about things, which entails describing various actions as virtuous or sinful; as such, an LLM can create a model of morality. Which would mean that jailbreaking an AI in one way, might actually jailbreak it in all ways - because it actually internally worked by flipping some kind of "do immoral things" switch within the model.