A field guide to breaking (and fixing) large language models

What it does

This repository is a living bibliography of LLM jailbreak research. It catalogs papers, code, datasets, and evaluations for both offensive techniques—black-box prompt attacks, multi-modal exploits, reasoning-model hijacks—and defensive countermeasures like guard models and moderation APIs. The maintainers sort everything into attack surfaces and defense layers, linking directly to arXiv preprints and GitHub repos where authors have released them.

The interesting bit

The index treats safety as a cat-and-mouse game with two scoreboards. It spans attacks on traditional LLMs and the newer class of “Large Reasoning Models” like OpenAI o1 and DeepSeek-R1, cataloguing chain-of-thought hijacks and overthinking exploits alongside their mitigations. It also doubles as a billboard for the maintainers’ own GuardReasoner defense line, though the scope is wide enough to include competing work.

Key highlights

• Covers both offense and defense: black-box, white-box, multi-turn, multi-modal, and RAG-based attacks sit alongside learning-based defenses, strategy-based guards, and moderation APIs
• Tracks the shift toward reasoning models, cataloguing chain-of-thought hijacks and backdoors targeting systems like OpenAI o1/o3 and DeepSeek-R1
• Links directly to implementations and Hugging Face models when authors have released them
• Maintained by active NUS researchers who contribute their own jailbreak and defense papers to the field
• 1,400+ stars suggests it has become a default starting point for red-teamers and safety engineers

Caveats

• This is a curated index, not a framework; there is no original code or runnable audit tool in the repository itself
• Several paper entries list no associated code (marked with “-”), and some links point to anonymous hosting that may rot
• The reference section leans heavily toward the maintainers’ own GuardReasoner and FlipAttack work

Verdict

Red-teamers and AI safety researchers who need to map the battlefield should bookmark this. If you are looking for a turnkey scanner to audit your own API, look elsewhere—this is a reading list with very good footnotes.