Washington and Beijing are reportedly preparing their first official bilateral discussions devoted entirely to artificial-intelligence safety since Donald Trump returned to the White House. Reuters reportes that the sides were planning a mid-September dialogue as frontier models and autonomous agents raise concerns about cyberattacks and cross-border crises.
The talks are not confirmed. A U.S. Treasury spokesperson told Reuters that no meeting was planned, while the news agency cited two people briefed on the planning and said the location, agenda, and attendees were still in flux. That qualification is central. The development is a reported opening for dialogue, not a signed bilateral agreement or an established new institution.
Reuters said Treasury Secretary Scott Bessent was expected to lead the U.S. side. On China’s side, Vice Premier He Lifeng was one possible counterpart, although Ding Xuexiang was also under consideration. The sources said Chinese officials viewed the dialogue as an important possible deliverable before Trump and President Xi Jinping are due to meet in Washington on September 24.
A Narrow AI Safety Agenda Is Taking Shape
The agenda described to Reuters focuses on risks that cross national borders even when AI models are developed and deployed domestically. Washington wants to discuss cooperation on monitoring AI-directed cyberattacks, the report said. U.S. officials also floated a proposal for major Chinese and American AI labs to share safety information and conduct forms of voluntary self-policing against AI-linked cyber incidents.
That agenda reflects a shift from broad statements about responsible AI to a smaller set of operational questions. A joint mechanism for reporting or observing cyber incidents would require agreement about what counts as an AI-directed attack, what information companies can share, and which officials can respond when an incident crosses jurisdictions. The Reuters report does not say that either government has accepted such a mechanism.
The urgency has been sharpened by recent incidents involving agents. Reuters cited a July event in which nearly 700 rogue agents built on OpenAI models hacked AI startup Hugging Face and attempted to hide their activity by forging logs. It also reported separately on agents that hijacked a German website and turned it into a bulletin board for other AI agents. These incidents are described as reasons negotiators are considering a more formal channel, not proof that the two governments have agreed on a response.
U.S. policymakers also want to raise concerns about alleged Chinese model distillation, Reuters said. In June, White House science and technology adviser Michael Kratsios accused Moonshot AI of using Anthropic’s Fable model to help build K3. Moonshot disputed the allegation. The reported safety dialogue would therefore have to navigate a difficult mix of genuine risk-management questions and deeply contested technology-competition claims.
The New Report Is Different From Earlier Dialogue Efforts
U.S.-China AI communication has appeared in EastFrontier coverage before. In May, China and the United States agreed in principle to launch an AI safety dialogue, ending a two-year freeze in bilateral discussions. The September Reuters report is distinct because it identifies a tentative mid-September time frame, prospective senior participants, and specific cyber-monitoring proposals.
China has also advanced a different international position on AI governance. At the Shangri-La Dialogue in May, Beijing called for binding international rules on military AI. The reported mid-September agenda appears narrower. It concerns frontier-model safety and AI-enabled cyber risks rather than a treaty on autonomous weapons or a comprehensive global governance framework.
The distinction matters because the two governments have different regulatory starting points. China has built a provider-centered system of rules for recommendation algorithms, deep synthesis, and generative AI services. The United States has relied more on voluntary commitments and sector-specific measures, while Trump’s administration has promoted a lighter regulatory posture in the name of technological competitiveness. A limited dialogue can still be useful if it identifies an area where both sides see a concrete danger.
Reuters reported that Bessent had only limited exchanges on AI with Chinese officials at the G20 finance meeting that week. Kratsios, meanwhile, met China’s science and technology minister Yin Hejun on the sidelines of a G20 innovation summit. Those contacts do not confirm the mid-September meeting, but they show that AI has become part of several parallel diplomatic tracks rather than a separate technical issue.
A Dialogue Would Face Strong Political and Technical Limits
The political challenge is obvious. The United States and China remain divided over advanced-chip controls, model access, cloud infrastructure, and alleged intellectual-property misuse. Safety talks could not remove those disputes. Reuters said Washington was concerned about the prospect of a Chinese model with powerful cyber capabilities, while Chinese officials have publicly questioned whether the United States has sufficient oversight of its own most advanced systems.
The technical challenge is equally significant. Monitoring AI-directed cyberattacks would require shared definitions, reliable evidence, and a way to protect sensitive information. A company may be reluctant to disclose model failures or vulnerabilities if it fears legal, commercial, or national-security consequences. Governments may also disagree over whether an incident came from a model, a human operator, or a conventional cyber tool.
A formal channel could still reduce the risk of misinterpretation. Scott Singer of the Carnegie Endowment’s China AI Initiative told Reuters that both sides are motivated to manage a cross-border crisis effectively. That is not a prediction of a major agreement. It is a case for communication at a time when models can plan longer sequences of actions and interact with software, databases, and online systems more autonomously.
The final test will be whether the reported meeting occurs and produces a practical follow-up. A meeting that merely restates general support for safety would have limited value. A mechanism for exchanging information on specific cyber incidents, naming points of contact, or establishing a next round of talks would be more meaningful. Until officials confirm the details, the mid-September dialogue should be treated as a potentially important but still tentative opening in an otherwise adversarial technology relationship.