US, China urged to cooperate on AI safety risks

Key takeaway

• AI researchers from the US and China gathered at a Beijing conference to warn that the two countries must cooperate on AI safety despite their competitive tensions, because increasingly capable AI models could enable cyberattacks and systemic failures that harm both nations.

• The conference highlighted universal cyber risks from advanced AI, including new vulnerabilities in AI-generated code and novel attack methods, and emphasized that collaboration on shared safety principles is crucial as open-weight models become more powerful.

3 Key Points

FAQ

What specific AI safety risks did the conference highlight?

The conference sessions focused on new kinds of vulnerabilities in AI-generated code, novel ways of attacking systems enabled by agentic AI tool use, and automated methods for carrying out social engineering attacks. One professor noted that hackers are expected to gain an advantage over the near term, though new AI-based countermeasures should eventually shift the balance toward defense.

Why are Chinese open-weight AI models a concern for US policymakers?

Chinese companies like Z.ai, Alibaba, and Moonshot have released highly capable open-weight models. As these models advance, it becomes harder to ensure they cannot help hackers find security vulnerabilities or be used as cyber weapons—and the latest model from Z.ai, GLM 5.2, includes frontier agentic and coding capabilities, according to expert analysis.

What actions has the US taken recently to restrict Chinese AI access?

The US government ordered Anthropic to prevent foreign nationals from accessing its most powerful models, Mythos and Fable 5, over national security concerns. In response, Anthropic revoked access for everyone. The US has also imposed tight restrictions on chips and chipmaking equipment sent to China to limit the country's AI development.