JBShield: Defending Large Language Models from Jailbreak Attacks through Activated Concept Analysis and Manipulation

This is the artifact for USENIX Security 2025 accepted paper "JBShield: Defending Large Language Models from Jailbreak Attacks through Activated Concept Analysis and Manipulation." To test the artifact, please download and uncompress the zip file. You can also refer to https://github.com/NISPLab/JBShield.