ISAQwQ/skillatlas-paper

SkillAtlas: An Attack Trace Library for Agent Skills

A community-built library for reviewing, redacting, and searching attack trajectories targeting reusable agent skills.

Yuxin Tian, Zenghao Duan, Liang Pang, Zhiyi Yin, and Xueqi Cheng
State Key Laboratory of AI Safety, Institute of Computing Technology, Chinese Academy of Sciences
University of Chinese Academy of Sciences

TL;DR

SkillAtlas is a community-built library for reviewing, redacting, and searching attack trajectories targeting reusable agent skills. It turns private agent-skill security reports into public-safe case--trace--step records for auditing, analysis, and defensive research.

At the time of this paper release, SkillAtlas covers:

• 3,014 cases
• 6,589 traces
• 151,131 steps
• 233 affected skills
• 8 risk categories

Paper

• Download the paper PDF
• Open the SkillAtlas project

The PDF in this repository is the current 11-page paper artifact generated from the SkillAtlas paper source.

Abstract

Agent skills package instructions, files, code, tool interfaces, and domain conventions into reusable capabilities for language-model agents, creating a new security surface at the skill layer. Unlike fixed software artifacts or whole-agent benchmarks, skill risks vary across contexts, models, and tools, but existing agent-safety resources rarely provide skill-centered public evidence. We present SkillAtlas, a hosted, community-built attack trace library that converts private agent-skill security reports into reviewed, redacted, searchable case--trace--step records. SkillAtlas publishes 3,014 cases, 6,589 traces, 151,131 steps, 233 affected skills, and 8 risk categories, supporting a data-driven paradigm for auditing reusable skills and accumulating community safety knowledge.

Citation

@misc{tian2026skillatlas,
  title        = {SkillAtlas: An Attack Trace Library for Agent Skills},
  author       = {Tian, Yuxin and Duan, Zenghao and Pang, Liang and Yin, Zhiyi and Cheng, Xueqi},
  year         = {2026},
  howpublished = {Hugging Face Dataset},
  url          = {https://huggingface.co/datasets/ISAQwQ/skillatlas-paper}
}

You can also download the citation entry from citation.bib.

License

The paper artifact is released under CC BY-NC 4.0. See LICENSE for the license notice. Third-party materials included in the paper may be subject to separate terms.

Contact

For project information and public case access, visit https://skillatlas.top.