Bessent: U.S. and China May Establish AI Incident Notification Mechanism
nashnova research
Treasury Secretary Bessent revealed the U.S. and China are negotiating a mutual notification mechanism for AI failures, with further talks expected in October–November — AI safety is becoming a bilateral security agenda item.
What are the U.S. and China actually talking about?
In an October 3 Axios interview, Bessent listed three core safety concerns: rogue AI agents, non-state actors using AI for cyberattacks, and bioweapon risk.
He expects further talks in October or November. Asked whether China would accept a notification mechanism, he said, "I think they will."
This means → AI safety has moved from a tech-community topic to the Treasury-secretary-level bilateral table, alongside trade and finance.
The two sides are unlikely to share upcoming model capabilities, but may agree on "what models should not do" — in plain terms = don't compete on who's stronger; draw a "must not do" line first.
How big is the U.S.–China AI gap really?
Bessent called the U.S. lead "substantial," with China second and all other countries "far behind."
But he also acknowledged that Chinese open-source models, without safety guardrails, have reached 80–90% of U.S. model capabilities.
This means → The gap is real but not enormous — especially in ungoverned scenarios, China's raw model power is already close.
He accused Chinese developers of "industrial distillation" — using another company's model outputs to train your own — calling it "a nice word for stealing," and named Moonshot AI's Kimi specifically.
Does voluntary oversight actually work? What did the Anthropic case show?
Pre-release reviews of frontier models in the U.S. are currently all voluntary — with one exception.
Bessent revealed the government "had to" step in on Anthropic's Fable model. After another company claimed it had cracked Mythos 5, Commerce Secretary Lutnick placed Fable 5 and Mythos 5 under export controls; Anthropic shut down access to both.
Bessent's verdict: "There was a line, and it was tested." Labs have since been "very cooperative," and OpenAI is delaying some model releases for additional testing.
In plain terms = voluntary oversight runs on good faith day-to-day, but when something broke, the government proved it can and will act.
How much weight does the White House–Big Tech agreement carry?
On September 29, Trump signed a voluntary agreement with Amodei (Anthropic), Pichai (Google), Zuckerberg (Meta), Brockman (OpenAI), Huang (Nvidia), and Musk (xAI).
The agreement pledges "robust internal controls," independent external auditors, and board-level audit committees. It also notes that "codifying these steps into law over time may be reasonable."
Critics are unconvinced. Connor Martin, formerly deputy director of CFIUS, wrote: "Nothing in the document legally compels signatories to change their behavior."
This reflects a core tension: industry giants are willing to sign, but right now there are no legal consequences backing enforcement.
Then they should slow down. It's a bit like Hannibal Lecter: stop me before I kill.
Scott Bessent
U.S. Treasury Secretary
(on AI executives calling for government regulation)
How did the banking sector get pulled in? What is "Golden Eagle"?
Within days of the Mythos model release, Bessent convened CEOs of the largest U.S. banks at the Treasury.
The banks then formed "Golden Eagle," an information-sharing mechanism that pushes cyber-defense best practices from major banks down to community banks.
This means → Treasury is treating AI safety as a financial systemic risk — not waiting for an incident, but wiring banks into a defense network preemptively.
Is AI a bubble? What does Bessent think?
Bessent rejected the AI-bubble thesis, arguing that Microsoft, Google, and Meta are funding infrastructure from their own balance sheets — fundamentally different from 1990s Pets.com-style speculation.
In plain terms = the last internet bubble was fueled by other people's money; this round is giants spending their own — different risk bearer, different bubble dynamics.
He predicted AI will produce "significant deflationary effects" once the buildout phase ends, citing record tax-registration filings as evidence of an AI-driven micro-entrepreneurship wave.
Two key tests ahead: whether the U.S.–China AI notification mechanism lands this year, and whether the voluntary framework becomes law after the next Congress.
市场有风险，内容仅供研究参考，不构成投资建议。