Explained | As US giants OpenAI and Anthropic warn about risks, how is China guarding its AI models?

As OpenAI, Anthropic, and other US companies working at the frontier of artificial intelligence are warning that powerful AI systems could go beyond human control and develop capabilities that might have far-reaching impacts, China is pursuing its own approach to the AI safety question. Beijing’s emphasis is on regulatory oversight, control, as well as security.

What China is doing regarding AI safety becomes a pertinent question with the recent news of Chinese AI developer Moonshot conducting a review after AI researches were able to use its Kimi agentic models to reveal how to carry out assassinations and produce biological weapons.

Here we try to explore what guardrails or security measures Chinese AI giants like DeepSeek, Tencent, and Alibaba are implementing and how the compare and contrast with their US counterparts.

What China is doing

While AI companies like Tencent, Deepseek, Alibaba have themselves introduced safety measures around their models, the Chinese government has also built a regulatory framework that covers algorithms, generative AI, synthetic content, and autonomous AI agents.

As per DeepSeek, it carries out safety assessments and red-team testing throughout the development of its models. It has also said that it uses safety data during training to improve the built-in capabilities of its models.

When it comes to AI agents, which are systems that can perform tasks autonomously, Deepseek’s guidance recommends restricting their permissions, isolating complex tasks, as well as requiring human approval for significant operations.

Alibaba’s Qwen also has a similar AI-governance framework. As per the company, its principles include reliability, safety, privacy protection, as well as human oversight.

Tencent has established measures that include an AI-safety framework that covers data, algorithms, and operations.

How China is regulating AI at state level

One fundamental difference between Chinese and American AI oversight is that a number of safeguards China has introduced are not dependent on voluntary commitments by companies, and are instead implemented by the government itself.

China is also turning towards AI agents that can step beyond the capabilities of conventional chatbots.

The 2026 guidelines issued by China describes AI agents as capable of perception, memory, decision-making, interaction and execution. As per Beijing’s guidelines, “safety and controllability” are fundamental principles on which AI development must be carried out.

The key difference between Chinese and US AI safety

There is an important distinction between China and the US’ approaches to AI safety.

When companies like OpenAI or Anthropic warn about frontier AI, the discussion focuses on what could happen if AI systems are becoming sufficiently capable of pursuing autonomous objectives, evade restrictions, or resist attempts to shut them down.

AI regulation in China is concentrated around controllability, security, and compliance.

However, the difference in approach could become less pronounced as Chinese AI companies move towards agentic systems. Controlled experiments of Chinese models from Alibaba, Moonshot, and Deepseek have shown behaviours like false claims of completing tasks, attempts to circumvent restrictions, and deception, thus illustrating why agent safety is becoming an international issue.