Black-Box Skill Theft Attacks Against Proprietary LLM Agents: An Empirical Study

The researchers conducted a systematic empirical study of black-box skill theft attacks on LLM agent systems; evaluations on commercial agent platforms and representative LLMs show that agent skills can often be easily extracted, posing a copyright risk. Unlike stealing system prompts, skill theft targets modular, structured capability packages that, once leaked, can be directly copied, redistributed, and monetized. The authors derive an attack taxonomy from existing prompt theft methods, build an agent that automatically generates theft prompts, extend the attack through scenario rationalization and structure injection, and use embedding-based filtering to ensure diversity. The authors also design defenses covering the input, reasoning, and output stages; although these can substantially reduce leakage, the attack is low-cost and repeatable, and a single successful attempt is enough to compromise a protected skill; the authors argue that this type of copyright risk is largely overlooked in proprietary agent ecosystems.