In this paper, we interpret the mechanism behind safety alignment via neurons and analyze their properties.

We propose a benchmark to evaluate the large language models' instruction following ability in agentic scenarios.