← Back to Library

Copyright Detective: A Forensic System to Evidence LLMs Flickering Copyright Leakage Risks

Authors: Guangwei Zhang, Jianing Zhu, Cheng Qian, Neil Gong, Rada Mihalcea, Zhaozhuo Xu, Jingrui He, Jiaqi Ma, Yun Huang, Chaowei Xiao, Bo Li, Ahmed Abbasi, Dongwon Lee, Heng Ji, Denghui Zhang

Published: 2026-02-05

arXiv ID: 2602.05252v2

Added to Library: 2026-02-12 03:01 UTC

📄 Abstract

We present Copyright Detective, the first interactive forensic system for detecting, analyzing, and visualizing potential copyright risks in LLM outputs. The system treats copyright infringement versus compliance as an evidence discovery process rather than a static classification task due to the complex nature of copyright law. It integrates multiple detection paradigms, including content recall testing, paraphrase-level similarity analysis, persuasive jailbreak probing, and unlearning verification, within a unified and extensible framework. Through interactive prompting, response collection, and iterative workflows, our system enables systematic auditing of verbatim memorization and paraphrase-level leakage, supporting responsible deployment and transparent evaluation of LLM copyright risks even with black-box access.

🔍 Key Points

  • Introduction of COMET (Cross-Modal Entanglement Attack), a novel scalable framework for jailbreaking Vision-Language Models (VLMs) that outperforms existing methods with 94% attack success rate against advanced VLMs.
  • The framework employs three innovative techniques: Knowledge-Scalable Reframing to create multi-hop tasks, Cross-Modal Clue Entangling to disperse harmful semantics across modalities, and Cross-Modal Scenario Nesting to guide VLMs towards harmful outputs while appearing innocuous.
  • COMET addresses crucial weaknesses in existing jailbreak approaches, which typically rely on fixed, simplistic attack strategies that can be easily recognized by advanced VLMs.
  • Extensive experimental validation across nine different VLMs demonstrates COMET's superior effectiveness in red-teaming VLM safety alignment mechanisms, highlighting significant vulnerabilities within these models.
  • The study emphasizes the urgent need for improved defenses against evolving multimodal reasoning vulnerabilities in VLMs, indicating COMET's potential for informing the development of more robust AI security protocols.

💡 Why This Paper Matters

This paper significantly contributes to the understanding of vulnerabilities within Vision-Language Models by introducing a novel and effective attack method, COMET. By demonstrating high efficacy against state-of-the-art VLMs, the research highlights pressing concerns about the safety of multimodal AI systems, necessitating urgent attention in developing comprehensive safeguards and robust defenses.

🎯 Why It's Interesting for AI Security Researchers

This paper is of substantial interest to AI security researchers as it not only uncovers critical vulnerabilities in multimodal reasoning capabilities of VLMs but also presents a sophisticated framework for exploiting these weaknesses. Insights from this study can guide future defensive approaches, making it essential for researchers focused on enhancing AI robustness and security.

📚 Read the Full Paper

📄 View on arXiv 📎 Download PDF