With tightening AI safety and ethical frameworks, how can enterprises build trusted AI systems to meet compliance requirements?

In September 2025, the "Artificial Intelligence Security Governance Framework" Version 2.0 was officially released, marking another significant step in China's AI regulatory landscape. Meanwhile, the EU's "Code of Conduct for General Artificial Intelligence" is set to take effect on August 2, ushering in a new era of "strong constraints" for global large-scale AI models. As of September 2025, China has released 30 national standards for artificial intelligence, with an additional 84 standards currently under development. Against this backdrop of intensifying regulatory measures, how enterprises can build a trustworthy AI system that both complies with regulations and maintains innovative vitality has become an urgent issue to address.

Building a trustworthy AI system requires systematic advancement across five dimensions, from foundational architecture security to continuous verification.

Foundational Architecture Security: Strengthening is required at every level from hardware to the application layer. Hardware selection must pass baseline security inspections, while the application layer should employ code vulnerability scanning and sensitive data masking techniques. For selecting a secure large model base, opt for base models that have passed authoritative security evaluations and utilize RAG technology to connect with enterprise private knowledge bases, ensuring output is precise and controllable.

Dynamic Defense Mechanisms: Establish risk awareness. Static defenses are no longer sufficient to address emerging AI risks. A multi-agent collaborative monitoring framework is essential. For instance, when an agent invokes a high-risk API, if the operation deviates from historical patterns, it should automatically trigger a request for manual confirmation.

Governance and Compliance Framework: The TRiSM governance framework comprises four pillars: explainability, model operations, privacy, and security. Model operations require implementing version control and rollback mechanisms to ensure quick recovery if performance degrades after updates. Privacy protection should employ dynamic data fencing technology, which automatically masks sensitive fields like patient IDs in healthcare scenarios.

Application Layer Security Reinforcement: Different application scenarios require tailored security measures. In government affairs scenarios, a cross-departmental collaborative defense network can be built to block lateral movement risks. In financial scenarios, integrating compliance components can ensure agents invoke pre-vetted APIs. Setting up manual intervention points is also crucial, such as automatically transferring emotionally charged customers from a customer service agent to a human representative.

Testing, Verification, and Continuous Evolution: Trustworthiness evaluation needs to integrate standardized testbeds covering assessments of various capabilities like communication, collaboration, and task execution. Regularly test using multiple types of attack algorithms to dynamically optimize defense strategies.

For example, iFlytek adopts an "AI for Public Welfare" approach, providing barrier-free services for hearing-impaired individuals. Hundsun Technologies is leading the development of industry standards for financial agents, promoting industry-academia-research collaboration and intellectual property protection. Alibaba Cloud, in partnership with the China Academy of Information and Communications Technology (CAICT), released the "Generative AI Excellence Architecture Design Guidelines," offering systematic architectural methodologies and best practices for enterprises of different sizes, helping them progress from "being able to use AI" to "using AI well."