Abstract: Publishing data that constitute trade secrets, such as unpublished manuscripts and proprietary reader data, constitute a core source of competitiveness for publishing enterprises in the digital-intelligence era. However, once AI agents are embedded into publishing operations, their model-training and content-generation processes may readily give rise to covert data leakage through circumstances such as unauthorized access and the incorporation of confidential information into prompts. AI agents may even be exploited by infringers as attack vectors for the improper acquisition of trade secrets and as auxiliary tools for circumventing or undermining confidentiality measures. The deep involvement of this technology not only amplifies data security risks at the factual level but also triggers a series of legal application dilemmas at the normative level, including disputes over the rights confirmation of generated content, difficulties in the determination of infringement and the challenges in allocating liability among multiple subjects. Facing the inevitable trend of digital-intelligent transformation, publishing enterprises urgently need to shift from a passive rights-defending mindset to systematically reconstruct a trade secret governance pathway that gives equal weight to prevention and relief from three dimensions: ex-ante privatized deployment and contractual constraints, in-process human-AI collaboration and privilege isolation, and ex-post evidence fixation and act preservation. This approach ultimately seeks to achieve a jurisprudential and practical balance between technological innovation and data security.