This repository is a continuously updated, curated collection of research papers on harmful fine-tuning attacks and defenses for Large Language Models (LLMs). It serves researchers, engineers, and AI safety practitioners by consolidating critical literature on LLM vulnerabilities introduced or exploited during fine-tuning, offering a centralized resource for understanding evolving threats and mitigation strategies.

How It Works

The project functions as a living bibliography, systematically cataloging and organizing academic papers. It categorizes research into distinct areas: attacks, defenses (spanning pre-training, alignment, fine-tuning, and post-fine-tuning stages), interpretability studies, and benchmarks. This structured approach facilitates a comprehensive overview of the threat landscape and the development of robust defense mechanisms.

Quick Start & Requirements

This is a curated list of research papers, not a software project. No installation or specific requirements are needed beyond accessing the GitHub repository. Links to papers and associated code are provided within the README for each entry.

Highlighted Details

• Comprehensive coverage of attacks and defenses, categorized by their stage in the LLM lifecycle.
• Regular updates (weekly) with recent research, including papers accepted to major conferences like NeurIPS, ICLR, and ICML.
• Highlights specific concerns such as jailbreak tuning, increased biorisk and cybersecurity risks, and vulnerabilities in large-scale models.
• Includes sections on interpretability studies and benchmarks for federated fine-tuning.

Maintenance & Community

The repository is actively maintained, with a last update noted as 2026/09/05. Researchers are encouraged to contact Tiansheng Huang (thuang374@gatech.edu) for missing references. The project also cites its own survey paper, "Harmful Fine-tuning Attacks and Defenses for Large Language Models: A Survey."

Licensing & Compatibility

No explicit software license is mentioned for the repository content itself. The listed research papers are subject to their respective publication licenses.

Limitations & Caveats

As a curated list, its primary limitation is reliance on the maintainers' ongoing effort to track and categorize new research. The scope is strictly limited to "harmful fine-tuning" and related attacks/defenses, excluding other LLM security vectors.