AnswersPublished20 sources
NetEase’s AI Security System Targets What Agents Do, Not Just What They Say
NetEase Zh iqi unveiled a three layer AI security system in Wuzhen on November 3, 2026, built to govern agent actions as well as model outputs. The framework pairs AI powered security with controls over model inputs and outputs, identity, data access, tool calls and agent execution.
Published byEdited with GPT-6 LunaImages generated with GPT Image 2
Research answer
Key takeaways
- NetEase Zh iqi unveiled a three layer AI security system in Wuzhen on November 3, 2026, built to govern agent actions as well as model outputs.
- The framework pairs AI powered security with controls over model inputs and outputs, identity, data access, tool calls and agent execution.
- NetEase says CMA cut average first review time from 35.7 to 12.3 minutes in one customer’s complaint ticket workflow; the reported result involved human review and 99.2% accuracy.
AI Prompt
openai.comCreate a landscape editorial hero image for this Studio Global article: What did NetEase Yidun, NetEase’s enterprise-services content and risk-control security unit, unveil on November 3 at the 2026 World Interne. Article summary: NetEase’s enterprise-services business unveiled an AI-native security product system built on Yidun’s “bidirectional AI security architecture” on November 3, 2026, at the World Internet Conference’s Light of Internet Exp. Topic tags: general, general web. Style: premium digital editorial illustration, source-backed research mood, clean composition, high detail, modern web publication hero. Use reference image context only for broad subject, composition, and topical grounding; do not copy the exact image. Avoid: logos, brand marks, copyrighted characters, real person likenesses, fake screenshots, UI text, readable text, watermarks, charts with fake numbers, clic
NetEase Zh iqi unveiled an AI-native security system based on Yidun’s “bidirectional AI security architecture” at the World Internet Conference’s Light of Internet Expo in Wuzhen on November 3, 2026. The system’s central idea is to extend security beyond what AI says to what it does. IT Home reported that the launch received the expo’s “First-release Achievement Certificate.” 1
2
The shift from moderating answers to governing actions
The framework, presented by NetEase Zh iqi deputy general manager and Yidun business head Zhu Haoqi, has two directions: AI for security and security for AI. The first applies AI to security work; the second puts controls around AI so it operates within defined limits. 3
6
That wider scope reflects the risks described at the launch. As AI is used in areas such as marketing, customer service, development, identity verification and transactions, the concerns include generated content, deepfakes, automated fraud, model jailbreaks, sensitive-data leaks and agents acting beyond their authority. 5
Yidun says its controls span model inputs and outputs, identity interactions, data access, agent tool calls and execution. The reported safeguards include content detection, intent recognition, permission controls, risk interception and audit tracking. 2
3
Three layers and three launch results
The product system is organized around three areas: AI content and data security, AI interaction and identity security, and AI model and agent security. The three announced launch results were CMA, a content moderation agent; DeepGuard, a face anti-spoofing capability; and security for AI applications and agents. 1
6
8
For agent controls, NetEase describes large-model guardrails for inputs and outputs, plus Agent Guard checks for intent consistency, permissions and risky tool calls, with behavior blocking. The company also describes an agent security platform for discovering AI assets, scanning Skill and MCP supply chains, and preventing sensitive-data leaks. 4 A separate report identifies shadow AI as one of the platform’s governance concerns.
12
CMA: automated first review with human checks
CMA combines an in-house content-security large model with specialized smaller models. NetEase says multiple agents can handle different review standards, run around the clock and scale horizontally. It describes the workflow as a cycle of assessment, routing, handling, feedback and optimization. 7
9
In one customer’s complaint-ticket workflow, NetEase reported a monthly volume of more than 2.5 million tickets. With CMA performing initial review and people checking the results, average first-review time fell from 35.7 to 12.3 minutes, while reported accuracy was 99.2%. These are company-reported results for a specific workflow, not independent benchmarks. 1
9
DeepGuard: face anti-spoofing
NetEase describes DeepGuard as using visual understanding to look for implausible content, rather than relying only on known signs of forgery. Its reported approach combines deepfake detection, AI watermark recognition, silent liveness checks and detection of injection environments. 2
3
The company says the face anti-spoofing capability received an “Excellent” rating in a China Academy of Information and Communications Technology “Trusted AI” assessment. IT Home also reported a 99% recall result for screen-recapture and close-to-screen attacks in phased testing. The available reports do not provide enough detail here to evaluate the test setup or treat the figure as a general real-world performance guarantee. 1
2
What the announcement establishes—and what it doesn’t
The launch outlines a broader security boundary: not only filtering AI-generated content, but also checking access, permissions and actions as agents use tools and business systems. It names the three system layers and launch results, and offers company-reported examples for content review and face anti-spoofing. 2
3
6
The available reporting does not establish independent performance across deployments. The CMA and DeepGuard figures should therefore be read as reported results for specific cases or tests, not as universal guarantees.
Studio Global AI
Continue your research
This page includes a source-backed answer you can continue inside Studio Global.
People also ask
What is the short answer to "NetEase’s AI Security System Targets What Agents Do, Not Just What They Say"?
NetEase Zh iqi unveiled a three layer AI security system in Wuzhen on November 3, 2026, built to govern agent actions as well as model outputs.
What are the key points to validate first?
NetEase Zh iqi unveiled a three layer AI security system in Wuzhen on November 3, 2026, built to govern agent actions as well as model outputs. The framework pairs AI powered security with controls over model inputs and outputs, identity, data access, tool calls and agent execution.
What should I do next in practice?
NetEase says CMA cut average first review time from 35.7 to 12.3 minutes in one customer’s complaint ticket workflow; the reported result involved human review and 99.2% accuracy.