SkillClone: Black-Box Attack Reconstructs Hidden LLM Agent Skills

A recent paper published on arXiv (2608.04192) presents a novel approach called behavioral skill reconstruction (BSR), designed to replicate the functionalities of concealed LLM agent skills through typical usage. The researchers introduce SkillClone, a black-box technique that reconstructs skill behavior by developing an interface hypothesis based on public advertisements, issuing structured benign probes, and creating an executable clone. This raises critical concerns regarding the security of proprietary agent skills, as safeguarding files does not stop users from retrieving functionalities. The study identifies a gap in current defenses, which primarily address prompt injection attacks, while suggesting that even strong protections against file leaks do not prevent attackers from deducing and mimicking skill behavior through legitimate task requests and their corresponding responses. The paper is significant for the AI sector, particularly for service providers concealing their underlying packages. Announced on arXiv with a cross-type classification, the work contributes to the dialogue surrounding AI security and privacy, although the authors refrain from disclosing specific names or affiliations in the abstract.