Chen Yixin’s AI Warning: Why China Sees Frontier Models as a Threat to Party Security

China’s top state-security official has placed AI squarely inside Beijing’s regime-security framework. Chen Yixin’s warning was broader than a conventional cyber-risk assessment: he portrayed advanced AI as a possible threat to the Chinese Communist Party’s political control, institutional stability, ideological security and critical information infrastructure.

That framing helps explain why Beijing can advocate AI governance while also treating domestic AI capacity, data access and advanced computing as strategic priorities.

AI as a political-security risk

Chen’s core concern was the potential use of generative AI at scale by adversaries. He warned that hostile actors could use AI-generated content to manufacture political rumors, spread harmful information and conduct public-opinion or cognitive warfare. He also linked AI to cyberattacks and to the collection of sensitive state data, trade secrets and personal information.

These are allegations and threat assessments by a Chinese security official, not public evidence that any specific foreign operation used a particular U.S. model. The important point is the breadth of the threat model: Beijing is treating information influence, intelligence collection, cyber operations and political stability as connected AI-security issues.

Why Claude Mythos and GPT-5.5-Cyber were named

Chen highlighted Anthropic’s Claude Mythos and OpenAI’s GPT-5.5-Cyber because advanced systems can assist cybersecurity work that is dual-use: the same capabilities that help defenders find, analyze and patch vulnerabilities can potentially lower barriers for malicious actors.

The underlying cyber capability is real and measurable, although evaluations do not establish that either model was used against China. The UK AI Security Institute reported that an early Claude Mythos Preview completed an end-to-end corporate-network attack simulation, while GPT-5.5 achieved a 71.4% average pass rate on the institute’s expert-level cyber tasks, compared with 68.6% for Mythos Preview. GPT-5.5 also completed the institute’s full simulation in two of 10 attempts. 1

Both companies have limited access to these high-capability systems. Anthropic says Mythos-class access is restricted to vetted organizations, while reporting on GPT-5.5-Cyber described an initial rollout to vetted cybersecurity professionals. 214

The policy response: build a “security barrier” around AI

Chen’s response was not simply to restrict foreign models. His approach called for a comprehensive security barrier around the AI stack, including:

• tighter oversight of domestic AI companies, algorithms and data;
• stronger protections for sensitive information and critical infrastructure;
• more secure domestic computing capacity and technology supply chains; and
• less exposure to foreign technological choke points, particularly advanced chips.

The logic is strategic self-reliance. In this view, AI security depends not only on model safeguards, but also on who controls the data, computing infrastructure and supply chain on which those models depend.

A dispute over what “AI safety” means

Chinese state-linked commentary has portrayed some U.S. calls to slow frontier-AI development as selectively applied: presented as safety measures, but viewed in Beijing as helping preserve U.S. technological advantage and reinforcing restrictions on China.

That narrative should not be confused with an objective assessment of model competition. Claims that Chinese open-source models are catching up on cost, performance or developer adoption depend heavily on the benchmark, deployment setting and model being compared. Cyber evaluations themselves show why broad rankings can mislead: performance differs by task, access conditions and the level of autonomy being tested. 1

Still, the political implication is clear. Beijing is unlikely to embrace international rules that it believes would constrain Chinese capability-building while leaving U.S. advantages in chips, cloud infrastructure and frontier models intact.

What this means for U.S.-China AI talks

The United States and China were reportedly preparing dedicated AI-safety discussions in mid-September, with frontier-model and AI-directed cyber risks among the potential topics. China viewed the dialogue as a possible deliverable for a September 24 leaders’ meeting, although a White House official said no AI meeting was then planned. 17

That creates room for narrow cooperation: communication channels on AI-enabled cyber incidents, exchanges on risk assessment, or voluntary practices for testing and access controls. It does not point to an agreement to pause frontier-AI development or to relax technology restrictions.

Chen’s warning captures the central divide. Washington and Beijing may share an interest in preventing AI-assisted cyber harm, but each government increasingly sees AI capability itself as a component of national power. For China, AI safety is therefore inseparable from political security and technological sovereignty.