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

Alan Tang, Lifeng Wang, Gang Wu, Tao Zhang, Yantao Jia, Xiaofeng Zhu

Huawei · Harbin Institute of Technology, Shenzhen

Published October 11, 2026 · on the feed October 15, 2026 · ▲ 15 · arXiv 2610.15725

The abstract

Agent skills encode operational procedures and domain expertise as reusable knowledge assets, yet their distribution in plaintext exposes them to unauthorized copying and leakage. However, practical mechanisms for skill ownership verification and leakage tracing remain underdeveloped. Watermark embedding before distribution offers a potential solution, but structured, compact skill instructions impose stringent requirements for factual and procedural accuracy. To address these challenges, we introduce SkillMarker, a KGW-based watermarking method that jointly targets functionality preservation and strong watermark embedding. SkillMarker unifies quality-constrained rewriting with signal-oriented candidate search. The former separates essential knowledge from flexible expression to preserve instruction fidelity; the latter explores alternative expressions to obtain a stronger watermark signal under preservation constraints. We release an open-source toolkit that integrates watermark embedding, quality auditing, and detection to support the protection of agent skills. SkillMarker targets ownership verification and leakage tracing in distributed skill ecosystems, where content may be copied in whole or in part, or incorporated into other skills.

The Brief

1:55

A short podcast summary of the paper. Written and voiced by AI.

Transcript

SkillMarker. A KGW-Based Watermarking Method for Agent Skill Ownership Verification. By Alan Tang and five others, from Huawei and the Harbin Institute of Technology, Shenzhen. Published on the eleventh of October. Its code repository is public but slight: one source file.

Agent skills encode operational procedures and domain expertise as reusable knowledge assets. They are distributed in plaintext, which exposes them to unauthorized copying and leakage, and practical ways to verify who owns a skill, or to trace where one leaked, remain underdeveloped. Embedding a watermark before distribution is a potential answer. The difficulty is the material itself. Skill instructions are structured and compact, and they must stay factually and procedurally accurate, so there is very little slack in the text to carry a mark without damaging what it tells an agent to do.

SkillMarker is a watermarking method based on KGW that aims at two targets together: preserving functionality, and embedding a strong watermark. It unifies two components. One is quality-constrained rewriting. The other is signal-oriented candidate search.

The rewriting separates essential knowledge from flexible expression, so the parts that carry instruction fidelity are preserved while the wording around them is free to change. The search then explores alternative expressions within those preservation constraints, looking for the ones that give a stronger watermark signal. One component guards what the skill says, and the other makes the most of how it can be said.

The work comes with an open-source toolkit that integrates watermark embedding, quality auditing, and detection, built to support the protection of agent skills.

The intended setting is a distributed skill ecosystem, where content may be copied in whole, copied in part, or incorporated into other skills. Ownership verification and leakage tracing are the two jobs the watermark is meant to do there. The tension SkillMarker is built around is that the mark has to live inside text whose every step must still work.

More with an account

Listen to every day’s papers

Daily playlists, deep dives, quests for the papers, and quizzes on their key terms.

Join free
KipQuest: the day’s AI research, as podcasts and quests. All papers Android app Privacy Contact Episodes are written and voiced by AI from each paper. Summaries can contain mistakes: the paper is the source.