Altman and Amodei ask the UN for AI standards. The White House says no
The OpenAI and Anthropic CEOs asked the Security Council for shared model tests, incident reporting and a bioweapons ban. The US delegate said dialogue cannot drift towards global governance.
Two lab CEOs asked for shared AI tests on September 23. The US delegate at the same table turned them down.
OpenAI chief executive Sam Altman and Anthropic chief executive Dario Amodei briefed the UN Security Council on September 23, 2026, at a high-level meeting France convened during General Assembly week. Both asked governments for shared rules on how frontier models get tested and how failures get reported. At the same meeting, White House science adviser Michael Kratsios said international dialogue "cannot be allowed to drift towards global governance", according to the UN's meeting record.
Yoshua Bengio, co-chair of the UN's scientific panel on AI, also briefed the council. His panel had already called this summer's agent incident an early warning, which we covered on September 22.
Two labs, two lists of mechanisms
Altman asked for common methods to measure AI capabilities and assess risk, meaningful human oversight, fast reporting of serious incidents, and secure channels between governments for sharing threat information, Implicator reported. He added a condition, quoted by PYMNTS: the standards "should not lock in incumbents or favor one business model over another."
Amodei went narrower. The UN record says he urged "narrow" agreements that can win broad support, with a ban on using AI to make biological weapons as his example. He also asked for evaluation and verification systems that keep pace with AI development, and common standards for testing models for loss-of-control and misuse risks. Implicator adds that he wanted systems that let countries verify each other's AI capabilities and commitments.
The US answer came from the same table
Kratsios, who directs the White House Office of Science and Technology Policy, also told the council that the US "totally rejects any attempt to construct a globalist scheme of control of superintelligence", according to the UN record. That line echoed President Trump, who called AI oversight a "globalist scheme" at the General Assembly on September 22, Fortune reported.
No written agreement came out of the session, according to Implicator.
A narrower incident channel, proposed to China
The US did put one incident mechanism on the table, just not a global one. Treasury Secretary Scott Bessent said the US proposed a notification mechanism with China for national security-related AI incidents, Fortune reports, after he met Vice Premier He Lifeng in New York that week. China's commerce ministry said on September 24 that the two countries had held their first dialogue on AI. Standing next to Trump at the White House that day, Xi Jinping said AI development must be "always under human control", per the same Fortune piece.
Fortune's report does not say what would count as an incident, who would send a notice, or whether labs would feed into it.
Why a build studio cares
Our read here is about engineering. The two labs asked for a shared way to measure capability, and the one government that hosts both of them said no to a global version. So for now, every benchmark in a model card is the lab's own number on the lab's own tests. If you ship on these models, the evals that match your product are the ones you write: a fixed set of your real tasks, rerun on every model version change before it reaches users. Incident reporting is in the same state. Both CEOs asked for it and nothing binding exists between labs and the teams building on their APIs. Decide now what you would log, what you would report to the model provider, and who on your side owns that call.
Next step: read the UN meeting record for each speaker's remarks, and our note on OpenAI's rules for outside testers for the lab-side version of shared evaluation. If you want a second set of evals on a model you already ship, write to us at hello@gattyworks.com.