This paper addresses the serious safety challenges posed by the increasing deployment of large-scale language models (LLMs) in real-world applications. Existing safety research primarily focuses on LLM outputs or specific safety tasks, limiting its ability to address broad and undefined risks. In this paper, we propose the Safe-SAIL framework, which leverages sparse autoencoders (SAEs) to extract rich and diverse safety-related features that clarify model behavior and effectively capture safety-related risk behaviors (e.g., generation of hazardous responses, violations of safety regulations). Safe-SAIL systematically identifies SAEs with the highest safety-concept-specific interpretability, describes safety-related neurons, and introduces efficient strategies to scale the interpretation process. The researchers plan to facilitate LLM safety research by publishing a comprehensive toolkit containing SAE checkpoints and human-readable neuron descriptions.
