yueliu1999/Awesome-Jailbreak-on-LLMs

Organizes attack and defense methods across multiple threat vectors—targeting reasoning models, black-box/white-box scenarios, multi-turn conversations, RAG systems, and multi-modal inputs—alongside guardrail approaches like learning-based defenses and guard models. The repository indexes implementation code and datasets alongside paper citations, enabling reproducible comparison of attack/defense effectiveness. Covers emerging safety challenges in reasoning-heavy LLMs (o1-style models) and multimodal systems alongside traditional text-based jailbreaks.

1,245 stars. Actively maintained with 15 commits in the last 30 days.