China’s AI safety agenda is moving toward a more explicit concern: advanced systems that could operate beyond effective human command. The latest signal came from State Security Minister Chen Yixin, who wrote that advanced U.S. models could threaten China’s critical information infrastructure. His intervention did not create a new law or binding regulation, but it elevated the loss-of-control question in a national security context and tied frontier AI risk to infrastructure protection.
That shift is not isolated. China placed a future AI loss-of-control scenario inside a 2024 safety framework, then expanded the theme in 2025 with a principle focused on trusted application and preventing loss of control. In May 2026, the Cyberspace Administration of China’s AI-agent guidance called on developers to discover, intervene in, block, and recover from improper agent behavior. Reuters framed Chen’s remarks within that evolving policy environment in its coverage of how China is preparing for the risk of AI escaping human control, highlighting how Chinese policy language has become more direct even without a new binding rule.
From infrastructure risk to loss of control
Chen’s intervention matters because it connects two policy concerns that increasingly overlap in Beijing’s AI governance debate: the security of critical information infrastructure and the possibility that advanced models could exceed human control. The emphasis is not only on familiar questions such as inaccurate outputs, harmful content, or misuse by bad actors. It also reaches toward a harder governance problem: whether advanced AI systems could create conditions in which human direction, operational boundaries, or security safeguards become insufficient.
The distinction is important. A senior political intervention can shape policy priorities without functioning as a statute or administrative rule. Chen’s remarks should not be treated as a new compliance regime. At the same time, they should not be dismissed as ordinary commentary. They place loss-of-control risk within the vocabulary of state security and infrastructure vulnerability, adding weight to a theme already present in China’s AI safety architecture.
The 2024 safety framework marked an earlier step by naming a future AI loss-of-control scenario. That move placed the issue within China’s formal safety planning without asserting that such an event had already occurred. It gave policymakers a category for a prospective risk and made control a matter for structured attention rather than speculative debate alone.
The 2025 expansion sharpened that framing by adding a trusted-application, preventing-loss-of-control principle. The grammar of that policy element is significant. Trusted application and the prevention of loss of control are linked, suggesting that acceptable AI use depends not only on performance or usefulness but also on the ability to keep systems within intended boundaries. In that framing, trust is inseparable from control.
This makes the current policy direction more specific than a general call for safe AI. It points toward a governance model in which controllability is part of the baseline expectation for advanced systems. The theme can apply to technical design, deployment procedures, monitoring duties, and emergency response planning, though the specific legal obligations depend on the relevant document. The central movement is conceptual: loss of control is becoming an explicit policy problem rather than an abstract future concern.
Agent guidance makes control a developer responsibility
The CAC’s May 2026 AI-agent guidance adds a more operational layer. It calls on developers to discover, intervene in, block, and recover from improper agent behavior. That sequence matters because it treats control as a continuing process rather than a one-time approval step. Developers are expected to identify problematic behavior, respond when it appears, prevent it from continuing where necessary, and restore normal operation after disruption.
The agent context is especially important. AI agents are often discussed as systems that can carry out tasks with some degree of autonomy. When policy language focuses on improper agent behavior, it moves the control debate closer to deployment conditions, where monitoring and intervention may need to happen dynamically. The CAC guidance does not amount to a comprehensive loss-of-control law. Its verified role is narrower: it addresses AI agents and describes developer-facing expectations for handling improper behavior. Still, its vocabulary fits the wider pattern in Chinese AI governance.
Reuters reported that Xi Jinping said in July that AI should remain under human control. That top-level formulation aligns with the direction of the 2024 framework, the 2025 principle, and the May 2026 agent guidance. These are separate policy signals, not a single legal event. Taken together, they show a consistent emphasis on human oversight and system controllability.
For companies and institutions watching China’s regulatory trajectory, the practical lesson is precision. China has identified future loss of control as a safety scenario. It has linked trusted application with prevention of loss of control. It has told AI-agent developers to handle improper behavior through discovery, intervention, blocking, and recovery. None of this proves that a loss-of-control incident has occurred, and it does not automatically convert every policy statement into a binding obligation. It does show that controllability is becoming a more prominent benchmark in the way Chinese authorities discuss advanced AI.
That benchmark could shape future compliance expectations even before a comprehensive new rule appears. Policy language often sets the terms for later standards, enforcement priorities, and institutional behavior. In this case, the terms are becoming clearer: systems should remain governable, human authority should remain meaningful, and developers should be prepared for improper behavior rather than assuming that safe deployment ends at launch.
Domestic priorities and the U.S.-China backdrop
Chen’s intervention can be read alongside EastFrontier’s earlier coverage of new calls for U.S.-China AI safety talks and the planned mid-September AI safety dialogue. Its distinct contribution is a state-security framing that identifies advanced U.S. models as a possible threat to critical information infrastructure.
That does not mean China’s AI safety agenda can be reduced to bilateral rivalry. The domestic sequence is broader and predates the latest intervention. The 2024 framework, the 2025 trusted-application principle, and the May 2026 CAC guidance all point to an internal governance concern: AI systems should remain controllable, and improper agent behavior should be addressed by developers. The U.S. dimension heightens the security framing, but the loss-of-control theme is already embedded in China’s policy vocabulary.
The careful reading is neither alarmist nor dismissive. Chen’s intervention is high-level and politically significant, but it is not a law. The 2024 and 2025 policy elements show a gradual formalization of loss-of-control language. The CAC’s agent guidance shows how control concerns can become more operational by assigning developers a role in detecting and responding to improper behavior. Xi’s July statement, as reported by Reuters, supplies a top-level political formulation that AI should remain under human control.
For policy readers, the central development is the narrowing of China’s AI safety language. Rather than treating safety only as a matter of content controls, reliability, or social risk, Chinese policy now more plainly names the possibility that advanced systems could move beyond effective human direction. That emphasis may influence future rules, technical standards, and diplomatic discussions. For now, it marks a sharper policy focus on control rather than a completed regulatory overhaul.