Scientists have introduced a new framework called SafeKey, which aims to reduce the risk of dangerous inferences from large language models by 9.6%. This innovative approach focuses on enhancing the safety of "key phrases" within model responses, thereby improving the overall security and reliability of model replies. The research team has identified two primary objectives for their work. First, they seek to improve the safety signals embedded in the initial content of key phrases. Second, they aim to refine the model's understanding of its own reasoning process to better detect potential safety threats. To achieve these goals, the team designed a Dual-Path Safety Head (DPSH) mechanism. This mechanism involves two prediction heads that operate during the training phase. One head assesses the full content of key phrases before they are generated, while the other evaluates the reasoning process behind the model's queries. By integrating these two pathways, the model can pre-emptively embed safety signals within its internal reasoning, making it more likely to trigger a "safety alert" when necessary. Additionally, the researchers proposed a Query-Mask Modeling (QMM) technique. This method masks all tokens in the input query, requiring the model to generate key phrases based solely on its own recapitulation and understanding of the query. The design of QMM ensures that the model must "believe" and "utilize" the internal reasoning that already carries safety signals, significantly enhancing the autonomy and stability of safety decisions. By combining these dual strategies, the SafeKey framework not only reduces the likelihood of dangerous outputs but also ensures that the model is more aware of the safety implications of its responses. This approach represents a significant step forward in the secure deployment of large language models, offering a robust solution to mitigate risks and improve trust in AI-generated content.