China Balances Open AI Security Risks, Innovation Strategy
- tech360.tv
- 9 hours ago
- 3 min read
China faces a regulatory challenge balancing open weight artificial intelligence models' security risks against a national innovation strategy. This strategy remains crucial to its technological rivalry with the United States, researchers indicate.
Open weight models allow public access to code for free download and local hardware operation. These systems traditionally lag proprietary models. Recent Chinese lab developments have narrowed this performance gap.
Zhipu AI's GLM-5.2 became the first Chinese large language model to rank among the top three globally. Users praised it as the country's first open weight model reliable for daily coding. And this signals a potential regulatory shift, analysts warn.
Mark Witzke, a scholar, stated China might soon find open weight models too dangerous for public release. This aligns with US concerns regarding cyber and biosecurity risks from advanced systems like Mythos.
Anthropic's Claude Mythos caused global industry alarm months ago. This powerful large language model identified and exploited cybersecurity vulnerabilities. Its advanced reasoning prompted US restrictions, fearing lower barriers for bioweapon development.
Washington showed similar caution recently. It imposed export controls on Fable 5, a public variant of Mythos redirecting sensitive queries. US authorities instructed Anthropic to block foreign access due to reported vulnerabilities.
Controls were removed after Anthropic committed to cooperating with the US government on frontier safety. But the incident displayed the considerable stakes of advanced artificial intelligence. China's timeline for a Mythos equivalent model varies.
The timeline for China to build an open weight model on par with Mythos is shrinking. Different projections exist, including an early period next year, or by an early month in 2027 based on compute budgets.
Michael Zeng Jinghan, a professor, noted the urgent need for governance as potent open source models have heightened this requirement. And open weight models are copied, fine tuned, and redistributed globally upon release.
This global distribution complicates traditional access controls. Zeng explained advanced open weight systems could ease malicious actors' operations. Their embedded safeguards can be removed once downloaded, posing a distinct risk.
The current uncertainty creates a substantial dilemma for Beijing. Mark Witzke noted open weight models are crucial to China's global diffusion strategy. Reversing this approach would sacrifice their primary AI soft power achievement.
Kristy Loke, a fellow at MATS Research, stated an outright ban by Beijing appears unlikely soon. She suggested more active, formalised regulatory responses to rising open weight risks are probable. So officials expect increased regulatory action.
Beijing holds increasing concerns regarding open weight AI related security risks, particularly within cyber and critical infrastructure sectors. Strengthening AI risk management systems is a goal in its current five year programme, Loke added.
Premier Li Qiang urged the government to ensure AI safety during a State Council meeting recently. He called for improved institutional rules in technology ethics and testing certification. Establishing a dynamic, adaptive, tiered safety regulatory framework was also advocated.
Zeng indicated Beijing's present position remains "ambivalent." He predicted the country will eventually implement targeted restrictions or graded access controls. This approach would resemble Washington's recent measures.
Restrictions might involve tighter rules for public distribution of advanced model weights. Limits could apply to using specific open weight systems within government agencies, state owned enterprises, critical infrastructure, or politically sensitive areas.
Regulators might also demand stricter registration, evaluation, watermarking, or security testing before models become available on Chinese platforms. And this aims to establish a more controlled release environment.
Chinese authorities, including the Ministry of Commerce, recently convened with Alibaba Group Holding, ByteDance, and Zhipu AI. Discussions centred on potentially restricting overseas access to China's most advanced AI models, according to a Reuters report. The involved companies provided no immediate comment.
China balances open weight AI security risks against its national innovation strategy.
Chinese open weight AI models are rapidly closing the performance gap with advanced proprietary models.
US regulators previously imposed export controls on advanced AI models due to security vulnerabilities.
Experts suggest Beijing may implement targeted restrictions or graded access controls for its own AI systems.
Chinese authorities have discussed potentially restricting overseas access to advanced AI models.
Source: SCMP
