AGENT SKILLS · TEXT WATERMARKING

Expertise, shared.
Origins, traceable.

SkillMarker embeds keyed watermarks into agent skills, supporting ownership verification and leakage tracing while preserving essential instructions.

Try the examples instantly. No sign-in or GPU required.

01 / EXPLORE THE WATERMARK

Small changes.
A measurable signal.

Compare two complete skills before and after watermarking. Inspect the wording and the recorded detection results.

Loading saved examples…

Bring your own skill.

Upload a complete SKILL.md and explore embedding and detection in the live app.

Open Hugging Face demo ↗

02 / THE METHOD

Protect the instructions.
Mark the expression.

Quality-constrained rewriting and signal-oriented candidate search adapt KGW watermarking to structured agent skills.

SkillMarker pipeline: quality-constrained rewriting, KGW-biased candidate search, reconstruction, and keyed detection.
The SkillMarker framework. Select the figure to view it at full resolution.
01

Constrain the rewrite

Separate essential content from flexible prose. Keep protected literals and structure in place.

02

Search for signal

Explore alternative expressions under a keyed logit bias and compare eligible candidates.

03

Reconstruct & inspect

Assemble the skill, audit its quality, and test suspect text with full-text or sliding-window detection.

Why watermark agent skills?

Skills package reusable expertise in plaintext. That same portability makes their instructions easy to copy, redistribute, or merge into other skills.

Motivation: reusable skills encode expertise; copying and merging create tracing challenges; keyed watermarks can provide evidence of origin.

03 / THE RESEARCH

SkillMarker: A KGW-Based Watermarking Method for Agent Skill Ownership Verification

Alan Tang1 · Lifeng Wang1,2 · Gang Wu1
Tao Zhang1 · Yantao Jia1,* · Xiaofeng Zhu1

1 Huawei Inc.
2 Harbin Institute of Technology, Shenzhen

* Corresponding author.

arXiv:2610.15725 · 2026

Cite this work

@misc{tang2026skillmarker,
  title = {SkillMarker: A KGW-Based Watermarking Method for Agent Skill Ownership Verification},
  author = {Alan Tang and Lifeng Wang and Gang Wu and Tao Zhang and Yantao Jia and Xiaofeng Zhu},
  year = {2026},
  eprint = {2610.15725},
  archivePrefix = {arXiv},
  primaryClass = {cs.CR},
  url = {https://arxiv.org/abs/2610.15725}
}