Dongwu Securities: AI agents expand security boundaries, and traditional cybersecurity is expected to usher in a new round of product upgrades

The Zhitong Finance App learned that Dongwu Securities released a research report saying that AI security is becoming a rigid and necessary investment in the intelligent transformation of enterprises, and application protection has opened up a new market for the cybersecurity industry. As the big model evolves from content generation to intelligent execution of business tasks, security requirements expand from input and output review to complete business process control, and the scope of protection extends from content compliance and sensitive information protection to cross-system authentication, operation isolation, abnormal suspension and recovery mechanisms, driving up the value of single customer procurement. AI brings two growth paths to the cyber security industry: new protection requirements and improvement of existing business efficiency. The commercialization of the industry will first be reflected in the sale of independent AI security products, and then expanded customer value through platform expansion, renewal, and continuous operation services. Commercialization will go through three stages of single-point protection, unified governance, and continuous operation. Manufacturers that can establish authority, audit, and continuous monitoring as standard products are expected to achieve continuous growth as customers' AI applications expand.

The main views of Dongwu Securities are as follows:

AI security is becoming a rigid and necessary investment in the intelligent transformation of enterprises, and application protection is opening up a new market for the cybersecurity industry

As large models evolve from content generation to agents perform business tasks, security requirements expand from input and output review to complete business process control. Early big models were mainly responsible for question and answer, search, and text generation. Model manufacturers could cover quite a few common risks through security training, interface restrictions, and content filtering; after agents were connected to databases, business accounts, and execution tools, the impact of misjudgment extended to enterprise production and data assets, and enterprises needed to independently set data access boundaries, operation permissions, and approval processes. Security requirements will gradually be deepened along with office assistance, business collaboration, and core production. The scope of protection will expand from content compliance and sensitive information protection to cross-system authentication, operation isolation, abnormal suspension and recovery mechanisms, and promote the increase in procurement value for individual customers. AI brings new protection requirements and two growth paths to the cybersecurity industry: products such as large model security gateways, permission management, and data leakage prevention directly generate revenue growth; AI empowers alarm investigation, vulnerability analysis and auxiliary response, reducing manual analysis costs and shortening disposal time. The commercialization of the industry will first be reflected in the sale of independent AI security products, and then expanded customer value through platform expansion, renewal, and continuous operation services.

The AI security industry chain presents a hierarchical collaboration pattern, and the enterprise deployment portal determines the competitive position of third-party security vendors in the market

Model vendors control weights, training processes and deployment strategies, leading security training, model alignment and risk capability assessment; third-party vendors focus on enterprise-side applications and infrastructure protection, combine general security capabilities with customer data boundaries, authority structures, and business processes, and professional evaluation agencies participate in independent verification. Enterprises simultaneously use external model interfaces, internal deployment models, and agents developed by different departments, which can easily cause inconsistent policies, scattered logs, and unclear boundaries of responsibility, further enhancing the value of unified security management across models. Third-party vendors form a product system covering access, execution, and audit through linked identity authentication, data governance, and security operation systems. Leading overseas model companies have carried out defense projects with security vendors to verify the industrial path of combining model capabilities and security delivery systems: after the vulnerability is discovered, it is still necessary to complete asset location, repair coordination, patch verification, and continuous tracking. The core barriers of traditional network security vendors come from data connections formed by long-term deployment, disposal authority granted by customers, and industry delivery experience. They can identify risks based on the business context and embed disposal in the production process.

Domestic AI security commercialization has already been implemented, and the realization of industry profits depends on revenue structure and delivery efficiency

Domestic manufacturers have disclosed an increase in AI security-related revenue, and demand for additional protection is being transformed into actual revenue contributions. There are differences in the statistical scope of each company. It is necessary to distinguish between revenue from independent AI products and revenue from traditional AI enabled products, and evaluate actual increases based on contract expansion, renewal, and delivery costs. Commercialization will go through three stages: single-point protection, unified governance, and continuous operation: initially procuring security gateways around specific applications, then expanding unified identity, policies, and auditing capabilities as the number of models and agents increases, and eventually forming continuous monitoring, attack testing, and security operation services. The higher the degree of standardization, the easier it is for manufacturers to reduce the cost of repeated delivery and increase the value of a single customer. The focus of industry products will extend from simple inspection to independent execution and control. Model protection is difficult to completely eliminate the risk of prompt injection. It should be combined with minimum authority, high-risk operation approval, and downstream system authorization and verification to limit the scope of influence after the inspection fails. Combined with domestic and foreign regulatory requirements for safety assessments, risk records, and incident reports, manufacturers that can establish authority, audit, and continuous monitoring as standard products are expected to achieve continuous growth as customers' AI applications expand.

Investment advice: Select targets around product launch, customer development, and profit restoration. Related companies include: Shenzhen Trust Service, Green League Technology, CITIC Security, Sino-Singapore SEC, Qianxin, Anheng Information, and Tianrongxin.

Risk warning: Enterprise AI deployment and security procurement fall short of expectations; demand for traditional cybersecurity continues to be pressured; product prices and gross margin decline due to increased competition; product protection effectiveness and reliability fall short of expectations; profit recovery and cash flow improvements fall short of expectations; risk of valuation fluctuations.