OpenAI and Anthropic urge UN AI safeguards after their chief executives warned the Security Council that advanced systems could become too powerful to control. The September 23 briefing put frontier-model risks directly before the body responsible for international peace and security.

 

OpenAI CEO Sam Altman and Anthropic CEO Dario Amodei asked governments to establish real controls, not merely voluntary promises. Their testimony also challenged diplomats to prevent artificial-intelligence capability from becoming concentrated in a single company or country.

 

OpenAI and Anthropic Press UN AI Safeguards

The two executives delivered unusually direct warnings during the high-level meeting in New York. According to Associated Press reporting, both argued that governments need safeguards capable of restraining systems before their capabilities move beyond effective human oversight.

 

The CEOs converged on three priorities:

  • Preserve meaningful human control over advanced AI systems.
  • Create enforceable safeguards across national borders.
  • Avoid concentrating AI power in one company or country.

 

The joint appearance matters because OpenAI and Anthropic compete for customers, researchers and computing capacity. Their public agreement does not settle how rules should work, but it gives governments a clearer industry signal that voluntary laboratory policies may not be sufficient for the most capable models.

 

Security Council Focuses on Human Control

The Security Council framed the discussion around international security rather than ordinary product regulation. Diplomats examined who should retain authority over life-and-death decisions as AI becomes more capable in cyber operations, intelligence analysis, military planning and autonomous systems.

 

That framing expands the policy question beyond misinformation, copyright or consumer protection. A model that can discover vulnerabilities, coordinate tools or act across networks can create cross-border consequences even when it was developed and deployed by a private company in one jurisdiction.

 

United Nations Secretary-General António Guterres had warned a day earlier about autonomous weapons and the danger of machines making lethal decisions. The concern is no longer limited to hypothetical superintelligence; it includes the present integration of AI into surveillance, targeting and cyber systems.

 

The Council can define threats to peace, impose sanctions and authorize collective measures, but it is not a global technology regulator. Any durable AI-control regime would still require member-state legislation, technical standards, independent evaluations and agreements that can be enforced across borders.

 

UN AI Governance Has Institutions but Limited Enforcement

The United Nations has already built a foundation for wider coordination. Its Global Digital Compact called for an international scientific panel on AI and a global dialogue that would give governments, researchers, industry and civil society a common forum for evidence and policy.

 

In a December 2024 Security Council debate, Guterres urged international guardrails, warned against delegating human oversight in weapons systems and asked states to advance the scientific panel and governance dialogue. The latest meeting shifts the debate toward what controls should bind frontier developers.

 

A scientific panel can help governments agree on capabilities and risks, but it cannot compel a laboratory to delay a model or disclose an incident. A dialogue can reduce policy fragmentation, yet it does not automatically produce inspections, reporting duties or penalties for violations.

 

The gap between coordination and enforcement is central. National rules can be undermined if developers relocate training, customers access models across borders or governments treat safety disclosures as strategic secrets. At the same time, an overly rigid global system could entrench large incumbents by making compliance unaffordable for smaller laboratories.

 

Diplomats Must Turn Warnings Into Verifiable Rules

The next policy challenge is translating broad agreement into measures that can be tested. Governments could require incident reporting, third-party evaluations, access controls for dangerous capabilities and documented shutdown procedures, with thresholds tied to what a model can do rather than the name of its developer.

 

Independent assessment will be especially important because the companies asking for regulation also define many of the benchmarks used to describe their systems. Regulators will need technical access and expertise sufficient to verify safety claims without forcing companies to disclose every trade secret publicly.

 

The Security Council meeting does not create a binding AI treaty or global regulator. It does, however, place loss of control and concentration of power on the international-security agenda with explicit support from two of the sector’s most influential executives.

 

Further progress will depend on whether governments can agree on measurable duties before another serious incident forces action. The practical test is not whether leaders accept that advanced AI can be dangerous, but whether they will establish rules that remain credible during commercial and geopolitical competition.

 

More on This Story