The 2026 World Internet Conference "Light of the Internet" Expo is being held in Wuzhen, Zhejiang, where NetEase (NTES) Smart Enterprise made its appearance and unveiled its first AI-native security product suite built on the Yidun bidirectional architecture, putting forward a key judgment: security must move from "controlling what AI says" to "controlling what AI does."
In response to this trend, Yidun proposed a bidirectional AI security architecture, with the core being: AI for security, and security for AI. AI for security means using AI to upgrade security capabilities. With the help of vertical models, multi-agent collaboration and multimodal understanding, AI can automatically analyze content and threat data, generate and verify security policies, allowing security personnel to shift from relying on manually writing rules to accepting and confirming operational results. Security for AI means using security capabilities to ensure controllable AI operations. Around large model input and output, identity interaction, data access, agent tool invocation and behavioral execution, it establishes content detection, intent recognition, permission control, risk interception and audit tracking capabilities, enabling AI to operate within clear boundaries.
Both directions share the same foundation, namely Yidun's long-term accumulation in content security, business risk control and security operations. On this basis, Yidun built a three-layer AI-native security product suite, which has been recognized as a "debut achievement" and awarded a certificate by the Light of the Internet release hall. CMA uses a self-developed content security large model combined with multiple specialized small models to build a review closed loop of "assessment—triaging—disposal—feedback—optimization," supporting 7×24 hour operation and rapid horizontal scaling. DeepGuard, based on visual understanding, moves from "looking for known forgery traces" to "understanding content and discovering unreasonable elements," and integrates Deepfake detection, AI watermark recognition, silent liveness and injection attack environment detection to make multimodal joint decisions. This capability passed the China Academy of Information and Communications Technology's "Trustworthy AI" face anti-counterfeiting capability specialized evaluation and was rated "Outstanding." The evaluation covered generative attacks, presentation attacks and injection attacks; in phased testing, the recall rate for screen re-photography and screen-pointing attacks reached 99% with 96% precision.
Yidun uses a large model security fence to guard input and output; uses Agent Guard to cover intent consistency verification, permission control, tool invocation risk detection and behavior interception; and uses an agent security platform to achieve asset discovery, Skill/MCP supply chain security detection and sensitive data leak prevention. In response to the widespread problems of "shadow AI" and invisible assets in enterprises, the platform provides unified management from two dimensions: AI asset security and AI operational security, and offers disposal capabilities such as interception, review, alerting and desensitization.
Public information shows that NetEase Smart Enterprise · Yidun is the trusted security foundation of NetEase Smart Enterprise. Founded in 2016, it has long focused on digital content security, intelligent risk control and AI security, serving more than 400,000 developers and thousands of paying customers. It is one of the drafting units of the national standard "Basic Security Requirements for Generative Artificial Intelligence Services."
Comments