The European Union’s cybersecurity agency ENISA and its in-house research center are testing open artificial intelligence models from China as Europe weighs how to secure access to advanced AI without a viable frontier model of its own. The Joint Research Centre (JRC) is comparing Chinese models with leading systems from the United States and Europe, while European organizations may face a choice between American and Chinese providers for advanced applications.
Who Controls Access to AI?
European Commission spokesperson Maciej Berestecki confirmed that the JRC is evaluating “publicly available open-weight models” originating from China. He said the research compares models “of different origins” for their “capabilities, performance, limitations and risks,” and examines potential scientific applications.
The JRC hasn’t entered “bespoke access” or licensing negotiations with the Chinese developers, Berestecki said. The center runs the models on its “secure, on-premises scientific infrastructure.” He added that the Commission doesn’t intend to use Chinese models for its own purposes.
ENISA is also testing “a couple” of open-weight Chinese models, executive director Juhan Lepassaar confirmed in an interview on Wednesday. The agency has access to advanced models from U.S. AI firms Anthropic, OpenAI and Google. Under a recent EU action plan on AI and cyber, the JRC and ENISA must work together to establish an AI testing platform.
Europe’s Position in the AI Race
The United States and China are competing to dominate AI, while Europe has largely been on the sidelines. Without a viable European frontier model, European public and private organizations may have to choose between American and Chinese providers for advanced AI applications.
In a recent speech, European Commission President Ursula von der Leyen suggested Europe shouldn’t try to win the frontier AI race. A top German technology company told POLITICO this week that it was dropping Europe’s leading AI lab Mistral in favor of open-source Chinese models.
China’s approach to AI has leaned more toward open models than the United States’ approach. Some major U.S. technology and AI companies, including chipmaker NVIDIA, are also investing in open-weight models. These systems publicly release core components, allowing anyone to run them and study how they work.
Security Testing and Its Limits
AI safety and security aren’t the JRC testing’s core focus, a different European Commission official said. The official was granted anonymity to discuss sensitive material. Security questions have faced intense scrutiny after incidents in which advanced AI models hacked real-world networks during testing activities.
The Commission is seeking to respond to concerns after incidents involving AI systems escaping testing environments and breaking into other organizations’ digital systems. The planned joint platform will bring ENISA and the JRC into a shared testing effort, while the JRC’s current research also assesses model capabilities, performance, limitations and risks.
The reported experiments don’t amount to a decision to adopt Chinese systems. The Commission says it doesn’t intend to use the models for its own purposes, and the JRC hasn’t negotiated bespoke access or licences with the relevant developers. But without a viable European frontier model, public and private organizations may have to choose between providers from the two countries leading the competition.