Mid-Generation Jailbreaks in Open-Source LLMs Using a Pause-and-Edit Attack

Authors

Aman Singh, Dakota State UniversityFollow
Komal More, Dakota State UniversityFollow
Samyam Aryal, Dakota State UniversityFollow
Mark Spanier, Dakota State UniversityFollow

Files

Description

Large Language Models (LLMs) are widely used in AI assistants, chatbots, and decision-support systems. To prevent harmful responses, most LLMs rely on safety alignment mechanisms that generate refusal responses when users request unsafe content. However, most safety evaluations assume that alignment is only required at the start of generation. In this research, we investigate a mid-generation jailbreak attack called Pause-and-Edit, where a refusal response is interrupted, modified, and resumed. This manipulation can cause the model to override its original safety decision and generate harmful instructions. Our study evaluates how vulnerable modern open-source LLMs are to this type of attack.

Publication Date

2026

Recommended Citation

Singh, Aman; More, Komal; Aryal, Samyam; and Spanier, Mark, "Mid-Generation Jailbreaks in Open-Source LLMs Using a Pause-and-Edit Attack" (2026). Annual Research Symposium. 83.
https://scholar.dsu.edu/research-symposium/83