The core assets of the AI industry are shifting toward model weights, which are the final products of model training. Because model weights are difficult to protect under copyright or patent law, their protection in practice relies on trade secret protection under Korea’s Unfair Competition Prevention Act. Building on the technical characteristics of model weights, this study examines each statutory requirement for trade secret protection. Model weights constitute technical information; in closed-source deployment, secrecy can be maintained despite the theoretical possibility of extraction; and large-scale model licensing transactions demonstrate their independent economic value. Regarding secrecy management, the most contested requirement in determining trade secret status, this study argues that technical, contractual, and organizational measures must be comprehensively established and that their adequacy should be assessed using objective recognizability as the essential criterion. Drawing on recent U.S. disputes involving model extraction and knowledge distillation, the study further argues that violations of terms of service constitute improper means only insofar as the violated provisions function as secrecy-management measures. It also proposes a four-factor framework for assessing extraction: the legitimacy of access, the mode of conduct, the purpose and result, and awareness of protective measures. Finally, the study argues that transparency obligations and trade secret protection can coexist through controlled disclosures designed according to the recipient, scope, and conditions of disclosure.