Black-Box Skill Stealing Attack from Proprietary LLM Agents: First Empirical Study
arXiv·high signal
Researchers demonstrate that high-value skills from proprietary LLM agents can be extracted via black-box observation. The study quantifies the emerging skill economy: free marketplaces report 90,368 published skills, paid marketplaces have 2,000+ listings and over $100K in creator earnings. The attack shows that skills — the reusable capability units that inject expert knowledge into general-purpose models — are vulnerable to extraction, threatening the economic model underpinning agent skill marketplaces.