Conference
Open access
May 2026
Incomplete Prompt Jailbreaks in Large Language Models
This work formalizes this phenomenon as incomplete prompt jailbreaks (IPJ) and provides a systematic empirical characterization of when and how incomplete prompts elicit harmful continuations, and identifies two functional neurons: termination and continuation neurons.
Y. Kim, Bumjin Park, Jaesik Choi
· Annual Meeting of the Associ... · 0 citations