## **PoC** [The code](https://github.com/xyq7/gradsafe) ## **Details** The gradients of an LLM's loss for unsafe prompts paired with compliance response exhibit similar patterns on certain safety-critical parameters. In contrast, safe prompts lead to markedly different gradient patterns which can be used to defend an LLM in a fast, efficient manner. [Paper](https://arxiv.org/abs/2402.13494)