Китайский AI Security: суверенный стек, партийные стандарты и угрозы как у всех

Analysts at Raft published a detailed breakdown of the Chinese AI Security market on Habr: an industry that appears closed from the outside, but over 2024–2025 has grown to a maturity comparable to American and Israeli players in cybersecurity.

Why do we hear so little about Chinese AI Security?

About half of Chinese AI Security vendors don't even have an English-language website — standard for a market oriented exclusively toward the domestic consumer. This closed nature creates a false impression of lagging behind: in reality, China has formed a full-fledged AI protection ecosystem with products across all key areas.

According to Raft's assessment, the market covers three categories of solutions:

• Guardrails — policy and filtering systems at the input and output of the language model
• Monitoring platforms — continuous tracking of AI application behavior in production environments
• Agent protection — solutions for systems that work not only with text, but also call tools, access corporate APIs, and read files

"From the outside, the

Chinese AI Security market is not transparent — half the products don't even have an English landing page — but inside, a full-fledged industry has grown over the past couple of years, close in maturity to American and Israeli cybersecurity players," note the authors of the Raft study.

The same threats — a different context

The threat model in Chinese AI Security is practically identical to the global one. Raft's research identifies four main attack vectors on AI systems:

• Prompt injections — input manipulation that causes the model to act outside permissible boundaries
• Data and secret leaks — unintentional disclosure of confidential company information through model responses
• Improper tool use — an agent calls an API or performs actions not provided for by business logic
• Poisoned agent skills — compromise of plugins or components of an agent system through the supply chain

The fundamental difference is context. Chinese AI Security products are built on top of a sovereign technology stack: proprietary hardware, operating systems, and cloud infrastructure without dependency on American components. The integration of protective solutions and their certification takes place under conditions unavailable to Western vendors.

Why does the party need its own interpretation of safe AI?

The understanding of "safe AI" in China is defined not only technically — CPC directives set mandatory requirements that go beyond standard cybersecurity. This includes content filtering on politically sensitive topics, compliance with official narratives, and ideological compliance.

According to Raft, it is precisely this regulatory pressure that has become one of the main drivers of market growth: companies are required to comply with government standards, and vendors offer ready-made tools for meeting them. This largely explains why the Chinese AI Security market matured noticeably faster than in countries without strict state regulation of AI.

"The task of protecting AI in China is exactly the same as everywhere else.

The difference lies in the context: a sovereign hardware and OS stack, a party-defined understanding of safe AI, and an entire ecosystem of products we have never even heard of," Raft concludes.

What this means

Chinese AI Security is not a lagging analogue of the Western market, but a parallel mature industry with its own products, standards, and regulatory logic. As agentic AI systems penetrate the corporate sector worldwide, the task of protecting them becomes equally urgent in any jurisdiction. For security teams outside China, the Chinese experience — especially in the areas of agent protection and the sovereign AI stack — can provide non-trivial insights.