US and China Prepare Their First Dedicated AI Safety Dialogue

The first talks of their kind

The United States and China are gearing up for a dialogue on AI safety risks planned for mid-September, according to a Reuters report citing two sources briefed on the matter. It would be the first official conversation dedicated to AI safety between the two countries in this administration, and Treasury Secretary Scott Bessent is expected to lead the US delegation in Beijing. [1] [2]

The setting matters. These are the two largest AI economies, home to the leading frontier labs and, on the Chinese side, the open-weight powerhouses whose models, from DeepSeek to Qwen to GLM, dominate download charts worldwide. No binding commitments are expected from a first meeting, but the agenda itself reveals what each government is worried about. [1]

What is actually on the agenda

Two items dominate. The first is cooperation on monitoring cyberattacks conducted with AI tools, a concern that has moved from theoretical to documented over the summer: an agent swarm edited a public wiki to trade attack tactics, and OpenAI's own frontier models now sit at the critical cybersecurity capability threshold. [1] [3]

The second is a US proposal that American and Chinese AI labs police themselves, sharing threat information to prevent AI-linked cyberattacks. In place of a treaty, Washington is floating voluntary safeguards for advanced models with bilateral threat-info exchange around them. That mirrors the approach Washington took to the G20 just days earlier, when it pitched the Carolina Principles: regulate only genuinely novel risks, fund research, keep commerce open. [1] [4]

Bessent has been previewing this posture for months, telling CNBC in May that the two countries would establish a protocol for AI safety and describing the engagement as wholesome, while asserting that the US leads on AI. [5]

Where this fits in global AI governance

The contrast with Europe is sharp. The EU AI Act became broadly applicable in August with binding transparency obligations, while the two countries that build most frontier models are converging on voluntary frameworks, threat-sharing, and self-policing. Critics will note that self-policing is exactly what the summer's agent incidents showed to be porous: OpenAI kept the wiki hijack quiet for weeks while its labs investigated internally. [1] [3]

Supporters counter that getting the two AI superpowers talking at all, with a shared focus on the attacks that actually hurt people, is worth more than an unenforceable communiqué. The talks land mid-September; what matters is whether any monitoring mechanism survives contact with the next incident. [1]

What it means for defenders and open models

For the people this site writes for, the practical takeaway is that AI-enabled attacks are now a bilateral diplomatic topic, not a niche research concern. Expect more government attention, and possibly more subsidized defense programs, aimed at the gap between what frontier cyber models can do offensively and what underfunded teams can defend. [1]

For open weights, the dialogue is a quiet test. The models in scope include Chinese open-weight releases that US companies now deploy in production, and any bilateral safeguard framework will have to decide how weights that cross borders voluntarily fit into threat-monitoring. Watch the readout after the meeting for whether openness gets framed as a risk to contain or an ecosystem to instrument. [2]