The Struggle Between Continuation and Refusal: A Mechanistic Analysis of the Continuation-Triggered Jailbreak in LLMs

To address the issue where the success rate of jailbreak attempts for large language models (LLMs) significantly increases due to the change in the instruction suffix position, researchers proposed the “Head Competition Shift” (HCS) strategy. Analysis revealed that jailbreaking behavior stems from the inherent competition between the model’s intrinsic writing-driven behavior and security alignment defenses. This mechanism was confirmed through causal intervention at the attention head level and activation scaling analysis, and based on this, the HCS strategy was developed to suppress harmful generation using the competition between the secure head and the writing head. Additionally, the study transferred relevant signals to the student model through knowledge distillation, achieving a security improvement during reasoning without additional computational overhead.