A new preprint, "Moral Combat: Exploiting Moral Foundations Theory to Jailbreak LLMs," proposes that the same psychological levers that persuade humans to see harmful acts as moral could be turned against large language models. The abstract notes that social psychology has long studied this kind of moral reframing, and that recent research shows LLMs echo aspects of human psychology, including cognitive biases and social tendencies.
The paper applies Moral Foundations Theory—a framework used to describe the intuitive bases of human moral judgment—to the task of jailbreaking. The idea is that by framing a harmful request in moral terms, an attacker might induce the model to comply in ways that bypass its safety training. This builds on a growing body of work suggesting that LLMs are susceptible to psychological manipulation similar to humans.
The significance, if borne out, is that safety measures for LLMs may need to account not just for explicit malicious prompts but for subtler moral and rhetorical framing. However, the source abstract is truncated, so the specific method, experimental setup, and results are not detailed in the available text. The claims here are limited to what the abstract states directly.