OpenAI chief Sam Altman and Anthropic chief Dario Amodei asked the UN Security Council to support common tests, reporting and controls for advanced AI.

The leaders of two frontier-model companies briefed the United Nations Security Council on 23 September about artificial-intelligence risks. Sam Altman, chief executive of OpenAI, and Dario Amodei, chief executive of Anthropic, warned that poorly controlled systems could threaten international security. Their appearance placed private AI laboratories inside a forum normally focused on state conflict and weapons.

Why it matters

Advanced AI is developed mainly by companies, but many of its risks cross borders. A model released in one country can support cyberattacks, biological research or influence operations elsewhere. Voluntary company rules cannot require rivals or states to report incidents. International standards could create shared evidence, even when governments disagree about broader regulation.

Associated Press reporting says Altman asked countries to align how they measure capabilities and disclose failures. Amodei argued for narrower agreements, including a prohibition on AI-assisted biological weapons, mutual evaluation and verification, common testing standards and an incident-notification system. These proposals are not treaties and the session did not create binding obligations.

The distinction between broad governance and narrow risk controls matters. Countries may resist an international body with authority over model development, yet still agree that a laboratory should report a dangerous escape, share a test method or prevent assistance with biological weapons. Comparable aviation and nuclear regimes began with specific technical practices before reaching deeper cooperation.

The strongest counterargument concerns incentives. OpenAI and Anthropic would help define standards that could shape their own market. Requirements expensive enough for smaller laboratories could protect incumbents. Company leaders also have commercial reasons to present their systems as powerful enough to require government attention. Independent technical institutions and transparent methods are therefore necessary.

What an agreement would need

A useful testing standard must specify what is measured, under which model settings and with which tools. A capability score without prompts, attempt limits and evaluator details is difficult to compare. Incident reporting likewise needs a threshold: minor product errors should not overwhelm a system intended for events with cross-border consequences.

Verification is harder than agreement. Governments may classify evidence about cyber or biological capabilities, while companies may protect model details as trade secrets. One practical design would let accredited evaluators inspect systems under confidentiality and publish limited findings, with stronger disclosure after serious incidents.

The Security Council also faces a legitimacy problem. Its five permanent members possess vetoes and include the largest geopolitical competitors in AI. Rules shaped only by major powers could exclude countries that experience harms without operating frontier laboratories. Any lasting regime would need broader UN participation or a separate technical body with global representation.

This meeting is evidence of agenda-setting, not completed governance. Readers should distinguish the leaders’ warning from an independently quantified probability of catastrophe. The public record confirms that they asked for coordination; it does not establish that their preferred mechanisms are sufficient or politically achievable.

The next concrete signal will be whether member states sponsor a resolution, commission a shared evaluation framework or establish a notification channel. Without an institution, timetable and reporting duty, the session remains a high-profile appeal.

Verification

  1. VERIFIED AS ASSOCIATED PRESS REPORTING — Sam Altman and Dario Amodei briefed the UN Security Council on 23 September 2026. Via: https://apnews.com/article/ai-artificial-intelligence-un-security-council-64519ea66b38e2600026f4481ad7f211
  2. VERIFIED AS REPORTED STATEMENTS — The leaders warned about loss-of-control and misuse risks. Via: the Associated Press report above.
  3. VERIFIED AS REPORTED PROPOSALS — Common tests, incident reporting and narrower biological-weapons controls were among the ideas described. Via: https://www.cnn.com/2026/09/23/tech/altman-amodei-ai-safety-un-security-council
  4. UNVERIFIED FROM A PRIMARY TRANSCRIPT — A complete official UN transcript was not located during this review. The article therefore attributes detailed proposals to reporting.
  5. VERIFIED — No binding international rule was created by the briefing itself. No resolution or adopted agreement was identified in the cited coverage.

Glossary candidates

  • Capability evaluation: A controlled test of what an AI system can do.
  • Incident notification: A requirement to report specified failures or harms.
  • Verification: A process for checking whether a party follows an agreement.

Cold-reader sentence: OpenAI and Anthropic asked the UN Security Council for shared AI tests and incident rules, but the meeting created no binding international safeguards.