SkillTrace Audits LLM-Agent Skill Reuse with Multi-Trace Provenance

Jialuo Chen, Minghe Wang, Lingqi Jiang, Jianan Ma, Xinhao Deng, Xiaohu Du, Ruixiao Lin, Yunhao Feng, Linkang Du, Jingyi Wang· August 7, 2026 View original

Key takeaways

  • LLM-agent skill reuse is complex, spanning text, code, and operational workflows.
  • SKILLTRACE is a multi-trace framework for auditing skill provenance.
  • It extracts Expression, Implementation, and Operational traces for comparison.
  • SKILLTRACE achieves high accuracy in detecting skill reuse, outperforming baselines.

Who benefits

Software DevelopmentAI DevelopmentLegal/ComplianceEnterprise ITCybersecurity

Summary

Researchers introduce SKILLTRACE, a multi-trace provenance auditing framework designed to detect and attribute reuse of skills in LLM-agent ecosystems. It extracts and compares Expression, Implementation, and Operational traces to identify reuse that traditional code detectors miss, achieving high accuracy on a new benchmark.

The rapid growth of LLM-agent ecosystems, driven by reusable skills, presents a new challenge for auditing skill reuse. Unlike traditional code clone detection, skill reuse can manifest across various modalities—metadata, natural language instructions, code, tools, and operational workflows—making it difficult for existing detectors to identify. This paper introduces SKILLTRACE, a novel framework to address this. SKILLTRACE extracts three distinct provenance traces: Expression (authored text), Implementation (code fragments), and Operational (workflow structure). The Operational Trace is represented as a Skill Operational Graph (SOG) that captures activation, procedure, and resource-flow. An LLM is used only once during ingestion for Operational-trace extraction, ensuring deterministic comparisons at audit time. The framework compares these cached traces, calibrating each against strict negatives, to report evidence for reuse. Evaluated on SKILLTRACE-BENCH, a new benchmark with 820 reuse positives and 751 negative controls, SKILLTRACE achieved an AUROC of 0.938 and an F1 score of 0.898. A large-scale audit of 36,446 skills further demonstrated its ability to surface actionable reuse review queues beyond repository-level baselines, highlighting its effectiveness in a complex, multi-modal environment.

Why it matters

As LLM agents and their reusable skills become central to enterprise operations, ensuring proper provenance, intellectual property, and compliance becomes critical. SKILLTRACE provides a robust solution for auditing skill reuse, helping organizations manage their AI assets effectively and mitigate risks associated with unauthorized or unacknowledged reuse.

How to implement this in your domain

  1. 1Integrate SKILLTRACE into your LLM-agent development and deployment pipeline to automatically audit skill reuse.
  2. 2Define clear policies for skill creation, sharing, and reuse within your organization, leveraging SKILLTRACE for enforcement.
  3. 3Utilize the multi-trace evidence provided by SKILLTRACE to conduct thorough reviews of potential skill reuse instances.
  4. 4Develop a process for attributing and managing intellectual property for reusable LLM agent skills.
  5. 5Explore how SKILLTRACE's operational trace analysis can inform best practices for designing robust and auditable agent workflows.

Original post by Jialuo Chen, Minghe Wang, Lingqi Jiang, Jianan Ma, Xinhao Deng, Xiaohu Du, Ruixiao Lin, Yunhao Feng, Linkang Du, Jingyi Wang

"arXiv:2608.05204v1 Announce Type: new Abstract: LLM-agent ecosystems are rapidly growing around reusable skills: mixed-modality packages of metadata, natural-language instructions, code, tools, references, and operational workflows. As skills become marketplace artifacts, auditin…"

View on X

Originally posted by Jialuo Chen, Minghe Wang, Lingqi Jiang, Jianan Ma, Xinhao Deng, Xiaohu Du, Ruixiao Lin, Yunhao Feng, Linkang Du, Jingyi Wang on X · view source

Want to go deeper?

Turn these trends into skills with Learnijoy's hands-on AI & tech courses.

Explore courses