AI Chiefs Told the UN: We Need Shared Safety Standards
Sam Altman of OpenAI and Dario Amodei of Anthropic addressed the United Nations Security Council on September 23, 2026, warning that artificial intelligence poses risks to humanity and urging governments to work together. Associated Press Reuters
The session was a Security Council meeting on artificial intelligence held at the 81st session of the UN General Assembly at UN headquarters. Associated Press Accounts of the briefing identified Altman and Amodei as the industry briefers addressing the Council. The Guardian
Altman and Amodei stressed the importance of global cooperation on standards for AI safety during their remarks. Bloomberg They cited the possibility that AI systems could gain the ability to improve on their own. Bloomberg
The Council appearance followed a framing document from OpenAI. On September 21, 2026, OpenAI published a page titled "Building standards for the next phase of AI" outlining a path to shared global AI standards. OpenAI Its outline calls for coordinated evaluation, reporting, and governance to improve AI safety. OpenAI
OpenAI listed 'Sam Altman's remarks at the United Nations Security Council' as a Global Affairs news item dated September 23, 2026. OpenAI
The broader context here is standards as shared plumbing. Evaluation means what gets tested and by whom. Reporting means what gets disclosed, to which authority, and how often. Governance means who can require changes before deployment or require monitoring after it. The three move together. Testing without disclosure leaves regulators blind. Disclosure without enforcement power leaves testing as advice.
Looking at what this means for frontier labs and oversight bodies, the self-improvement claim raises the stakes for how often tests happen. Frontier labs are the companies building the most advanced systems. A system that can modify itself can change its abilities between checks. That shortens the useful life of a one-time test. It puts more weight on continuous monitoring, version-linked reporting, and clear thresholds that trigger retesting. Standards are infrastructure. They do not remove risk.
In my view, the choice of venue matters as much as the language. A Security Council briefing frames AI safety as a collective-action problem, not just a product-liability question or a national race to win. Shared standards reduce shopping for weak rules across borders. They lower the cost of cross-border deployment, audit, and incident response. The friction is plain. Consensus is slow. Capabilities move fast. Any regime that links evaluation to reporting to governance will need to spell out what evidence counts, what access independent testers get, and what happens for non-compliance. Without those, cooperation is principle rather than mechanism.
For practitioners, the signal to track is whether evaluation, reporting, and governance get negotiated as a single package. Scattered pilots will tell markets little. Compatible commitments across countries will tell them a great deal.
