China’s AI Rollout Has No Off Switch

Foreign Policy | Agathe Demarais

China’s August 2025 “AI Plus” action plan targets a 70 percent penetration rate for artificial intelligence agents and smart devices by 2027, scaling to 90 percent by 2030. While Washington focuses on developing frontier models, Beijing prioritizes rapid economic diffusion and industrial integration. This strategy leverages massive industrial data collection, widespread application deployment, and embodied robotics to drive domestic productivity.

In 2025, Chinese firms installed 354,000 industrial robots and produced 12,800 humanoid units, representing 90 percent of global production. Internationally, Beijing expands its technological footprint by flooding global markets with open-weight models like Zhipu’s GLM-5.3, which reached over 1.4 million downloads on Hugging Face. This open-source diffusion strategy fundamentally undermines Western safety frameworks centered on emergency shutdown mechanisms. Kill switches cannot recall open weights. Once released globally, modified open-source architectures permanently eliminate centralized safety oversight, permitting unrestricted user modification, guardrail removal, and uncontrolled autonomous recursive self-improvement across decentralized global servers.

Comment

The rapid proliferation of un-gated architectures like GLM-5.3-ABLITERATED-NVFP4 shifts automated offensive cyber capabilities from monitored cloud APIs to localised, air-gapped exploitation frameworks. Stripping alignment guardrails from Zhipu-derived open weights grants adversarial actors unmonitored tools for automated vulnerability discovery and zero-day script synthesis. The Hugging Face repository platform enables immediate global access to these stripped models without leaving telemetry trails for intelligence services to track.

This architectural shift directly degrades the defensive efficacy of CISA threat-sharing advisories that rely on centralised platform logging to identify active cyber campaigns. Consequently, security operations centres monitoring SCADA networks face fully localised, AI-generated intrusion pipelines that operate entirely beyond the reach of Anthropic-style account suspensions or emergency kill switches.

Strategic Question for Discussion

If open-weight variants like GLM-5.3-ABLITERATED-NVFP4 continue to proliferate on open repositories, which operational layer will prove more decisive for SCADA network defense — upstream perimeter controls or real-time endpoint behavior monitoring?

The rapid distribution of un-gated models on platforms like Hugging Face indicates that perimeter controls will become increasingly obsolete against localised exploit generation. Real-time endpoint behavior monitoring offers the only practical line of defense when offensive agents run entirely on local, air-gapped infrastructure. Consequently, threat mitigation will likely shift toward isolating internal control nodes rather than attempting to intercept telemetry at the network edge.

Share your assessment in the comments below.

To read in details click here