Huawei's SkillMarker watermarks agent skills, detecting copied passages of 1,536 tokens
Alan Tang · hf · 2026-10-15
- Problem: agent skills distributed in plaintext are prone to unauthorized copying and leakage, yet structured skill instructions demand strict factual/procedural accuracy, making watermarking hard.
- Method: SkillMarker, a KGW-based watermarking method, combines quality-constrained rewriting (separating essential knowledge from flexible expression) with signal-oriented candidate search to strengthen watermark signals while preserving instruction fidelity.
- Results: negligible impact on downstream task performance across four datasets; window-based detection consistently identifies copied passages with as few as 1,536 consecutive tokens; detection rates of 72.2-77.8% persist after merging with one or three background skills.
- Open source: a toolkit covering watermark embedding, quality auditing and detection; project page at fuxi-ailabs.github.io/SkillMarker.
More from coding & agent
- awesome-llm-apps hits 142k stars with 100+ open-source AI agents, skills and RAG apps — Saboo_Shubham_ · 2026-10-15
- Dev Builds a 'Video Prompt Director' App on Eazo to Turn Rough Ideas Into Polished Video Prompts — alifcoder · 2026-10-15
- Dev builds a working animated Excalidraw with Gemini 4 Argon, plans to open-source it — Saboo_Shubham_ · 2026-10-15
- Layering Claude Skills over deterministic Makefiles for complex AI-automated workflows — bradneuberg · 2026-10-15
- Coding on a Walk: Dev Uses AirPods Voice to Run Codex Sessions, Ship Apps and Answer Email Hands-Free — nickbaumann_ · 2026-10-15
- Sentry founder denies removing telemetry options 'for enterprise dollars': everything is opt-in — zeeg · 2026-10-15