U.S. Deploys AI “Kill Switch” in June, China Explores Similar Measures for July
Beijing has engaged in quiet discussions with its leading artificial intelligence companies over the past month regarding restrictions on access to their advanced models, according to Reuters. Chinese authorities reportedly met with representatives from Alibaba, ByteDance, and the startup Z.ai to explore limiting overseas access to the nation’s most sophisticated AI systems, including those not yet released.
The meetings, convened by China’s Ministry of Commerce, involved discussions on imposing limits on both closed-source and open-weight models, the latter being accessible for local modification by developers. Officials also reportedly raised the possibility of classifying unauthorized disclosure or theft of proprietary AI technology as an offense under China’s national security law. Separately, proposals were floated to restrict the types of investors eligible to fund domestic AI startups.
The precise scope of any potential restrictions remains under debate, with some sources suggesting measures might only apply to future models rather than existing ones. No definitive timeline has been established, and the implementation of any new regulations is uncertain.
While the practicalities of such restrictions are unclear, a summary published in a Supreme People’s Court journal from a May roundtable on open-source AI regulation offered potential insights. Chinese legal experts proposed a three-tier system: basic open-source tools would require a simple government filing, more advanced technologies would undergo security reviews before release, and the most sensitive frontier models would be barred from public release or restricted to domestic use.
This proposed structure would represent a significant shift for Chinese AI companies, whose global growth has largely stemmed from open access. Alibaba’s Qwen series, for instance, has gained considerable traction on Hugging Face, a major repository for open-source AI models. ByteDance’s Doubao is a prominent AI product within China, and Z.ai’s GLM-5.2 has garnered attention from U.S. researchers for its performance comparable to top American models, offered at a significantly lower API access cost.
Any decision to restrict overseas access could increase costs for businesses that have come to rely on Chinese models as more affordable and less restricted alternatives to U.S. frontier systems. Concerns have also been amplified by the possibility that advanced cybersecurity models, such as Anthropic’s Mythos—which the U.S. administration restricted in June—could be reverse-engineered and used against Chinese systems, adding a defensive urgency to the discussions.
On June 12, Anthropic received a directive from the U.S. Commerce Department’s Bureau of Industry and Security to suspend all access to its Claude Fable 5 and Mythos 5 models for foreign nationals, including its non-citizen employees. Due to the technical challenges of selectively restricting API access, Anthropic temporarily withdrew both models globally. This marked the first instance of the U.S. applying export controls to a deployed AI model rather than the underlying hardware. The models were reinstated on June 30 after Anthropic updated its safety classifiers and the Commerce Department lifted the restrictions.
A similar pattern unfolded four days prior with OpenAI, Anthropic’s chief competitor. The company released its GPT-5.6 Sol, Terra, and Luna models, disclosing that it had previewed them with the U.S. government and, at Washington’s request, initially provided access to approximately 20 vetted partners. OpenAI has stated that government-gated access “shouldn’t become the long-term default.” President Trump’s executive order on AI issued in June had already encouraged developers to voluntarily submit frontier models for federal cybersecurity review prior to public release. A framework defining “covered frontier models” and the conditions for government pre-release access is expected by August 1.
Beijing has reportedly been closely monitoring these developments. Officials expressed alarm that Anthropic’s Mythos cybersecurity model could be weaponized against Chinese infrastructure to exploit software vulnerabilities. Additionally, concerns have been raised in China regarding allegations that Anthropic may be employing spyware-like tactics to track Chinese users.
Qihoo 360 founder Zhou Hongyi publicly voiced these concerns at ISC.AI 2026 in Beijing, advocating for China to develop domestic equivalents and unveiling Tulong Feng, a homegrown AI vulnerability agent.
The tiered structure proposed in China’s Ministry of Commerce discussions mirrors the U.S. approach to chip export controls over the past three years, a policy that has arguably narrowed rather than widened the technological gap with China.
Beijing has been progressively tightening its regulatory perimeter. In April, its state planning agency ordered Meta to divest its $2 billion acquisition of AI startup Manus, citing national security concerns under China’s foreign investment review mechanism. Moonshot AI and StepFun were required to obtain government approval before accepting U.S. capital in funding rounds. Furthermore, a broader regulatory package introduced in early June extended scrutiny to cross-border transactions involving Chinese technology and data.
The prevailing logic since the viral success of DeepSeek R1 in early 2025 has been that U.S. restrictions on frontier AI create a natural market for Chinese open-weight models. Chinese open-weight models saw a dramatic surge in usage on OpenRouter, a key global AI distribution hub, rising from less than 2% in late 2024 to approximately 61% by mid-2026. The U.S. defensive posture, therefore, inadvertently provided Beijing with a global distribution advantage that was not solely based on technical superiority.
This dynamic, however, hinges on the continued openness of Chinese models. If Beijing restricts overseas access to its frontier systems, whether closed-source or open-weight, this advantage could diminish. Z.ai’s GLM-5.2, for example, built its value proposition on borderless access under an MIT license; any restriction on this distribution would undermine its core appeal.
An analysis of the June U.S. controls highlighted a structural challenge that also applies to Beijing: national AI restrictions often struggle to remain purely national. The U.S. order on Anthropic, for instance, encompassed foreign nationals within the United States, including Anthropic’s own non-citizen employees. Data from MacroPolo indicates that over two-thirds of top-tier AI researchers in the U.S. received their training abroad, with foreign-born individuals comprising up to 70% in leading labs. Nationality-based access controls aimed at adversaries can inadvertently exclude the very engineers needed to address the vulnerabilities that prompted the restrictions. China faces a similar challenge in reverse. ByteDance and Alibaba are reportedly accelerating the rollout of humanlike agent features ahead of Chinese AI regulations scheduled to take effect on July 15, demonstrating Beijing’s capacity to implement restrictions on a defined schedule, independent of market feedback.
This potential shift could pose challenges for smaller labs and developers worldwide who rely on open-source technologies, an area where China has been a leading contributor.
French President Macron warned at the G7 summit that European governments would cease purchasing U.S. AI products if access could be abruptly curtailed. Canadian Prime Minister Carney described concentrated AI dependence as a “strategic mistake.” Both reactions implicitly assumed that Chinese AI would serve as an unconstrained alternative.
Leave a Reply