Generative AI presents three core legal challenges: the Collingridge dilemma (early intervention vs. regulatory uncertainty), Knightian uncertainty in risk assessment, and regulatory latency. The training of large language models raises critical trade secret concerns, particularly regarding the legal status of reverse engineering. While rights holders commonly impose contractual prohibitions on reverse engineering, enforcing such restrictions faces both theoretical and practical difficulties. Current legal frameworks lack clear standards for distinguishing between legitimate interoperability efforts and unlawful appropriation of protected elements. This paper proposes a dual approach: (1) establishing qualitative-quantitative criteria for trade secret protection, and (2) developing context-sensitive evaluations of reverse engineering methods based on transformative purpose, proportionality, and public benefit. The framework seeks to balance innovation incentives with intellectual property protection in China's evolving AI regulatory landscape, while clarifying the distinction between permitted decompilation and prohibited model extraction.