This paper investigates the vulnerability of safety-aligned large language models (LLMs) to jailbreaking attacks by exploring the concept of "safety patterns" within the representation space of LLMs. The study reveals that specific activation patterns in the representation space of LLMs act as "keys" that determine the model's ability to resist malicious inputs. These safety patterns can be identified using a simple method involving contrastive query pairs and are crucial for the model's defense mechanism. The research demonstrates that weakening or strengthening these safety patterns can significantly affect the model's robustness against jailbreaking attacks. Through extensive experiments, the study confirms that safety patterns are essential for LLMs to resist malicious queries, and their manipulation can lead to successful jailbreaking. The findings contribute to a deeper understanding of jailbreaking phenomena and highlight the need for the LLM community to address the potential misuse of open-source LLMs. The paper also presents a method for extracting safety patterns from LLMs and validates their effectiveness through various experiments and analyses. The results show that safety patterns play a critical role in the defense mechanisms of LLMs, and their manipulation can be used to either weaken or enhance the model's resistance to jailbreaking attacks. The study provides a new perspective on understanding and defending against jailbreaking attacks in LLMs.This paper investigates the vulnerability of safety-aligned large language models (LLMs) to jailbreaking attacks by exploring the concept of "safety patterns" within the representation space of LLMs. The study reveals that specific activation patterns in the representation space of LLMs act as "keys" that determine the model's ability to resist malicious inputs. These safety patterns can be identified using a simple method involving contrastive query pairs and are crucial for the model's defense mechanism. The research demonstrates that weakening or strengthening these safety patterns can significantly affect the model's robustness against jailbreaking attacks. Through extensive experiments, the study confirms that safety patterns are essential for LLMs to resist malicious queries, and their manipulation can lead to successful jailbreaking. The findings contribute to a deeper understanding of jailbreaking phenomena and highlight the need for the LLM community to address the potential misuse of open-source LLMs. The paper also presents a method for extracting safety patterns from LLMs and validates their effectiveness through various experiments and analyses. The results show that safety patterns play a critical role in the defense mechanisms of LLMs, and their manipulation can be used to either weaken or enhance the model's resistance to jailbreaking attacks. The study provides a new perspective on understanding and defending against jailbreaking attacks in LLMs.