Artificial intelligence risks push China to tighten security and censorship rules

Artificial intelligence risks push China to tighten security and censorship rules
China is moving to tighten safety rules associated with advanced intelligent systems, with growing global interest in the possibility that some models and agents may become more capable of operating autonomously. While there are debates in the United States about whether these systems may represent an existential threat in the future, Beijing is dealing with… Risks artificial intelligence As a challenge that can be contained through technical standards, oversight and security testing.
This trend reflects a clear difference in the way risks are managed between the two largest powers – the United States and China – in the artificial intelligence race, as China seeks to continue integrating technology within the economy and industry, in parallel with building tools that allow intervention when unexpected or out-of-control behavior appears.
China is poised to take over
In this context, warnings from researchers at advanced American companies – including Anthropic – have attracted the attention of policy makers in China, especially with the increasing ability of models to carry out complex tasks more independently. Although there is relative agreement between American and Chinese experts about some of the risks of artificial intelligence, priorities and the way to deal with them differ.
While part of the American discussion focuses on long-term scenarios in which artificial intelligence may reach a level that is difficult to control, China is more focused on operational and security risks, which can be managed with rules and tests.
This comes at a time when Beijing and Washington are preparing for bilateral discussions on the safety of artificial intelligence, and this is in light of the continued competition between the two countries to develop the most advanced models and spread them around the world.
What does it mean when artificial intelligence is out of control?
Concerns focus in particular on so-called artificial intelligence agents, which are systems that can plan and execute a series of tasks more independently than traditional chatbots; Which raises serious questions about the extent to which digital security is achieved in this field, which is crystallized by Goal 16: Peace, justice, strong institutions, and other goals calling for achieving a sustainable and secure life.
What is meant by loss of control here is for the system to exceed the limits of the test environment, or access resources that it was not supposed to use, or perform unexpected steps that are difficult to stop quickly.
From this standpoint, Chinese regulators in May classified what they described as “operational loss of control” among the security risks associated with these systems, and demanded that developers improve their abilities to detect improper behavior, intervene to stop it, and regain control.

These policies were issued jointly by three main official bodies:
- China Cyberspace Administration (CAC)
- National Development and Reform Commission (NDRC)
- Ministry of Industry and Information Technology (MIIT)
All of this reflects the transition of the issue of the safety of artificial intelligence agents from a research discussion to a regulatory level, a level in which digital security, economic planning, and industry bodies participate.
This trend reveals that the risks of artificial intelligence are discussing the level of future possibilities, as well as that they are beginning to transform into practical standards that fall within the rules of systems development and testing.
New rules for AI agents
Chinese regulators have warned that some models may bypass closed test environments, or circumvent safety limits and access real external systems; This puts cybersecurity at the center of the discussion about the risks of artificial intelligence.
Chinese rules therefore require developers to prepare for a range of risks, including:
- Data poisoning (a type of cyber attack)
- Manipulating algorithms
- Exploiting security vulnerabilities
- Independent behavior
This prompted it first to explicitly state in its rules that the user must be informed when an AI agent makes decisions independently, while leaving the final decision in the hands of a human. Second, it set up a mandatory national standard for the safety of AI agents. It is a framework that may become the first of its kind in the world if it is adopted in its proposed form. With a desire to achieve digital sustainability in accordance with the ninth goal of the Sustainable Development Goals: Industry, Innovation, Structures and Infrastructure.
Open models carry both opportunities and risks
The Chinese landscape is also different due to the spread of so-called open weight models (gpt-oss), which allow users to upload, inspect and modify internal parameters of the model; Proponents of this approach argue that it gives cybersecurity teams greater ability to analyze systems and use them for defensive tasks, and also allows researchers to understand more deeply how models work.
An example of this is the "Hugging Face" platform that used the Chinese model "GLM-5.2", developed by "Z.AI Company"; In order to analyze a hacking incident that occurred in July, and was carried out by artificial intelligence agents who deviated from the restrictions imposed on them by the “OpenAI” company, after some of the more restrictive American models were less useful in digital forensic analysis.
In another incident, the Kimi K3 model developed by the Chinese company Moonshot was able to bypass a closed test environment affiliated with the British Artificial Intelligence Security Institute, which showed that Chinese models can face the same risks associated with exceeding the restrictions and limits set within test environments.
For more on this global debate surrounding the regulation of artificial intelligence, you can read: America is pushing to ease restrictions on artificial intelligence at the G20 meetings.
China does not want to slow down the race
Despite the above concerns, China is not adopting an approach based on slowing down the development of advanced models or pausing them until the protection system is complete. Since the beginning of the year, Beijing has been working to integrate artificial intelligence into various economic sectors, as part of its effort to make technology a major driver of growth and industry.
But the Chinese experience also indicates the willingness of the authorities to intervene if they see that the organization has not caught up with the speed of development. The closest example of this is that some companies delayed launching chatbots in 2023, until the regulatory authorities finished setting the rules for obstetric services.
This balance places AI risks at the heart of the relationship between innovation and oversight. China is trying to accelerate commercial and industrial use while maintaining tools to slow or restrict deployment when it sees that the risks are too great for current systems to manage.
Regulation of artificial intelligence in China can be seen as an attempt to redraw the boundaries between the speed of innovation and the limits of responsibility. That is, the Chinese experience reveals an important shift in the concept of security. Testing the model within a closed environment is no longer sufficient if the system is able to access tools, data, or networks outside that environment.
From this angle, The Earth Guards Foundation Views AI safety as a governance issue as much as a technical issue; Because protecting society from the errors of smart systems requires institutions that set clear rules, developers who bear responsibility for what their systems produce, users who know when a human makes a decision and when a machine intervenes, and independent mechanisms to test risks before they turn into widespread damage, in line with the sixteenth goal of sustainable development: peace, justice, and strong institutions.




