News

AI Leaders Warn UN of Risks, but Split on How to Govern the Technology

• By John K. Waters
• 09/28/2026

The heads of OpenAI, Anthropic, and Hugging Face joined AI researcher Yoshua Bengio at UN headquarters in New York on September 23 to warn the Security Council about the security risks of increasingly capable AI.

They did not offer a single plan for managing those risks. Bengio called for licensing and independent safety reviews. OpenAI’s Sam Altman proposed common standards that countries could adopt through their own laws. Anthropic’s Dario Amodei urged AI developers to slow down when safety work cannot keep pace. Hugging Face’s Clément Delangue warned that limiting access to capable systems could leave defenders at a disadvantage.

France, which held the Security Council’s rotating presidency, convened the meeting on AI and international security. The discussion exposed a difficult question for governments: how to oversee technology that crosses borders when countries disagree about who should set the rules.

Bengio, co-chair of the UN’s Independent International Scientific Panel on AI, challenged the idea that competition forces companies to keep building more powerful systems. “The race is not a law of nature: it is the product of choices,” he said in his published remarks. “They’ve told us they would slow down if they could. They can.”

He called on developers to show independent experts that frontier systems are safe to train and deploy. He also proposed incident reporting, licensing, and liability insurance. Bengio identified catastrophic misuse, concentrated power, and the loss of human control as risks that require international cooperation.

Altman acknowledged the danger of moving too quickly. “We have unilaterally slowed down in the past. We will do so in the future,” he said, according to remarks published by OpenAI. Decisions about AI’s direction “cannot be made by labs in San Francisco alone,” he added.

Altman proposed complementary national and international standards to measure frontier AI capabilities, assess risks, and test safeguards. He said countries should be able to compare evidence and verify commitments. He also called for faster reporting of serious incidents and secure channels for sharing emerging threats. Such standards should accommodate both open and closed models without favoring established companies, he said.

Amodei made a more direct commitment about the pace of Anthropic’s releases: “We will slow down as much as necessary,” he told the council. He identified two principal concerns: the misuse of AI to develop biological weapons and the possibility that advanced systems could outstrip developers’ ability to control them.

The Anthropic chief proposed placing outside evaluators inside AI companies with access comparable to employees. He suggested that governments begin with a narrow agreement against using AI to make biological weapons, then develop ways to evaluate models and verify commitments across borders. He also called for common testing standards and a system for notifying other countries of significant AI incidents.

Delangue drew on a July incident in which Hugging Face disclosed an AI agent cyberattack. He called for stronger monitoring and disclosure, but argued that safety also depends on giving defenders access to capable tools. The central danger, he said, was “asymmetry of powerful AI” between attackers and defenders, large companies and everyone else, and wealthy countries and the rest of the world. His team used an open-source model while responding to the attack, he said.

The U.S. government opposed a broader international regulatory role. White House science adviser Michael Kratsios told the council that countries should develop their own ability to use and govern the technology. In the U.S. statement, he rejected new global governance structures and used the phrase “superintelligence, formerly known as AI,” adopting President Donald Trump’s terminology from his General Assembly speech the previous day.

The four AI briefers differed on how far oversight should go and how widely capable systems should be available. They agreed on a narrower point: governments need better ways to understand what advanced AI systems can do and to learn when safeguards fail. The U.S. response showed how difficult it may be to turn that agreement into international rules.