China Weighs Export Controls on AI Models Over Cyber Risks

China's Commerce Ministry is exploring new export controls on domestically developed AI models, according to a July 7 Reuters report, signaling Beijing's growing unease about the cybersecurity capabilities its AI sector has produced.

The ministry has convened meetings with major Chinese AI companies including Alibaba, ByteDance, and Z.ai to discuss potential restrictions on both closed and open-source models. The deliberations reflect a fundamental tension in China's AI strategy: balancing its preference for open-source diffusion with concerns about advanced systems falling into the wrong hands.

Proposed Control Framework

The controls under consideration would establish a tiered approach. Lower-capability models would face basic licensing requirements, while higher-end products would be subject to extensive monitoring and control regimes. The ministry is also weighing stricter criminal penalties for model distillation—a form of intellectual property theft where competitors reverse-engineer proprietary systems. U.S. frontier labs have previously accused Chinese firms of conducting such attacks.

These discussions come as Chinese authorities investigate Manus, a domestic AI startup with ties to Meta, for potential export control violations related to attempts to relocate operations outside China.

Parallel U.S. Struggles

The United States has faced similar challenges in regulating its most advanced AI systems. In early June, the Trump administration issued an executive order requesting voluntary 30-day pre-release testing periods for AI models to assess cybersecurity vulnerabilities. That framework quickly unraveled when Anthropic released its Fable model without submitting to the testing process.

The Commerce Department initially responded by imposing sweeping export controls on Anthropic's Mythos and Fable models, barring all foreign nationals—including the company's own employees—from accessing them. Those restrictions were lifted after three weeks of consultations, but the episode illustrated the difficulty of balancing innovation with security concerns.

Why it matters

Both superpowers are discovering that controlling AI diffusion is far harder than controlling chips or other physical technology. China's consideration of export controls on open-source models—which are inherently designed for wide distribution—highlights a strategic contradiction that could force Beijing to choose between global AI influence and domestic security. Meanwhile, inconsistent U.S. policy has created uncertainty that may be driving American companies toward Chinese alternatives, undermining the very security objectives these controls aim to achieve.

Competing Capabilities

Chinese firms have made notable progress in developing cybersecurity-focused AI systems. Z.ai has publicly deployed products with capabilities similar to Anthropic's Mythos, which can rapidly identify software vulnerabilities. Chinese open-source models have gained traction in U.S. markets due to lower costs, with Microsoft reportedly considering offering a domestically hosted version of DeepSeek.

The Foundation for Defense of Democracies, which first reported these details, recommends that U.S. agencies develop standardized approaches for sharing advanced AI capabilities with allies rather than imposing sudden restrictions that erode trust. The organization also emphasizes continued restrictions on China's access to cutting-edge chip manufacturing equipment as a more effective control point.

These details were first reported by the Foundation for Defense of Democracies based on Reuters reporting.