As large models flood into critical sectors such as finance, government, and energy, the "the smarter, the more dangerous" security paradox is raising industry awareness. At the 2025 Wuzhen Internet Conference, 360 Digital Security Group officially released the first domestic "Large Model Security White Paper," systematically revealing five core risks throughout the entire lifecycle of large models, and for the first time proposing a dual governance framework of "external security + platform-native security," establishing a layered defense barrier for the high-risk AI era.
The white paper points out that the security threats of large models have formed a complex matrix across multiple levels and with strong coupling, covering five dimensions:
Infrastructure layer: Attacks on computing clusters and training platforms may lead to model poisoning or theft;
Content layer: The generation of false information, prohibited content, or "AI hallucinations" may cause social risks;
Data and knowledge base layer: Leaks of training data or contamination of knowledge bases will undermine the trust foundation of models;
Agent layer: If AI agents capable of autonomous decision-making are controlled, they may perform malicious tasks;
User end layer: Attacks such as prompt injection and unauthorized access can bypass protection to directly manipulate model behavior.
Facing this multi-dimensional threat, 360 proposes a dual-track security strategy:
"External Security": Deploying monitoring, filtering, and auditing systems outside the model, such as content compliance gateways and abnormal behavior detection;
"Platform-Native Security": Embedding security capabilities within all stages of large model development, training, deployment, and inference, achieving "security as code."
Based on this concept, 360 has built a full-chain solution covering seven core capabilities including data anonymization, model hardening, content filtering, Agent behavior auditing, API protection, red-blue team exercises, and security compliance, which has been implemented in multiple industries such as finance, manufacturing, and government.
360 emphasizes that enterprise-level protection alone is not enough. In the future, it will work with academia, industry, and research institutions to promote the development of large model security standards, sharing of threat intelligence, and co-building of open-source security tools, creating an open, collaborative, and trustworthy AI security ecosystem.
AIbase believes that at this crucial turning point where large models move from "technological showcases" to "production infrastructure," 360's white paper is not only a risk warning but also a roadmap for building the new infrastructure of AI security. When intelligence becomes productivity, security must become the baseline — this defensive battle initiated by 360 concerns not only technology, but also the future of whether AI can be truly trusted and entrusted by society.
