AI safety goes before the Security Council
Altman and other AI leaders are due at the UN. The test is whether safety proposals produce access, evidence and accountability.
Morning edition · September 23, 11:00 a.m. Eastern. The UN Security Council is scheduled to hear from AI leaders this afternoon. The question worth watching is what governments will require companies to demonstrate before increasingly autonomous systems do more consequential work.
OpenAI’s Sam Altman, Anthropic’s Dario Amodei, Hugging Face’s Clément Delangue and researcher Yoshua Bengio are expected to brief the council. France is convening the discussion on AI and international security. Security Council Report’s agenda preview describes a focus on evaluation, verification and the risk of losing human control. Bloomberg reports that Amodei is scheduled to participate remotely. This is a preview of the meeting, not a report of an agreement.
The political disagreement is already clear. Reuters reports that President Trump used his September 22 UN address to oppose international control of AI and favor continued growth, while saying the Justice Department would watch the industry. The same report says the U.S. and China have discussed notification arrangements for AI incidents reaching a national-security threshold ahead of Thursday’s leaders’ meeting. A communication channel could help countries respond to trouble; it would not itself settle how to prevent it.
OpenAI has put a specific proposal on the table. Its September 21 standards paper calls for shared capability measurements, human oversight of automated AI research, and common incident classifications and reporting thresholds. It explicitly says these technical standards would not constitute licenses or mandatory prerelease approval. Governments would decide whether to incorporate them into law. That distinction matters: agreement on how to measure risk does not automatically give an independent body authority to stop an unsafe release.
There is also a useful example of what an evaluation can—and cannot—establish. In its September 22 assessment of Claude Opus 5.5, METR describes incremental improvements on its AI-research tasks and judges full automation of AI research unlikely for that model. It does not assess the model’s alignment properties or verify compliance with a particular safety threshold. METR discloses an unpaid agreement, limited testing time and Anthropic’s opportunity to review the published text. Those boundaries are part of the finding, not fine print to discard.
Our test for today’s discussion is concrete: who gets access to evaluate the systems, what evidence becomes public, and who can require a change when safeguards fail? Speeches and shared terminology can begin that work. They are not evidence that it is finished.
For readers trying to make sense of the debate, start with the methods and limitations behind a claim. Our Safety & Society guide separates practical precautions, documented harms, policy arguments and frontier research so that useful tools and serious scrutiny can sit together.