Altman speaks at the Council's 10228th meeting

On 23 September 2026 OpenAI published Sam Altman's remarks at the United Nations Security Council, labeled remarks as delivered. The same day, UN News covered the Council briefing under the headline that OpenAI and Anthropic briefed the Security Council amid a "real and imminent" runaway-AI threat framing, and noted Turing Award recipient Yoshua Bengio among the briefers. UN Web TV lists the session as the Security Council 10228th meeting on artificial intelligence and international security.

Altman's text frames AI as either a Renaissance-like expansion of human agency or an Industrial Revolution-scale disruption. He argues labs must refuse systems they cannot keep under human control, rejects concentrating the most powerful models in one company or country, and asks for complementary national and international frontier AI standards.

A briefing binds nobody

A Security Council briefing and a company remarks page do not enact treaty text.

  • UN member states on the Council decide whether follow-on resolutions, presidential statements, or informal processes pick up standards, incident reporting, or information-sharing asks.
  • National regulators and CAISI-class institutes decide whether to treat OpenAI's asks as input to evaluation, audit, and RSI-measurement programs.
  • Frontier labs (OpenAI and peers named in the UN coverage) remain bound by their home-law duties and by any voluntary frameworks they already signed. The remarks do not create a common license regime by themselves.
  • Critical-infrastructure operators are named as intended recipients of secure vulnerability channels. No mandatory enrollment mechanism is specified in the OpenAI post.

Two failure modes and four asks

  • OpenAI posts a full remarks-as-delivered transcript tied to a dated Council appearance. Operators can quote Altman's words directly from that primary.
  • Altman states agreement with Bengio on catastrophe pathways, then names two failure modes he wants avoided: loss of human control (including via recursive self-improvement), and concentration of power in too few hands.
  • He says OpenAI has "unilaterally slowed down in the past" and "will do so in the future," and that competitive pace is not a reason to accept excessive technological risk.
  • He lists three operating principles: people at the center of AI decision-making; scientific and economic gains by and for people; and individual empowerment. On the third, he says OpenAI is "not the heroes of this story" and that its role is to enable people worldwide.
  • Concrete international asks: shared standards for measuring capabilities, assessing risks, judging safeguards, and preserving human oversight; comparable evidence and verification language; speedy incident reporting and classification; secure channels among governments, critical infrastructure, and technical experts.
  • He cites OpenAI's recent math progress narrative, including a claim that a model recently "solved one of the Millennium Prize Problems, the Navier-Stokes equations." That claim sits in the remarks; independent verification is outside this Dispatch.

No resolution and no self-improvement test

  • No Council resolution, annex, or binding standard is attached to the OpenAI page.
  • "Slow down" is a corporate posture statement. It does not define triggers, external auditors, or publish criteria for pausing training.
  • RSI is named as a governance priority. The remarks do not ship a public RSI benchmark or measurement protocol.
  • Open vs closed model parity is asserted as a design goal for standards. Implementation detail is absent.
  • Anthropic and other briefers may have delivered different asks. This piece tracks the OpenAI primary; it does not reconstruct every speaker from wires alone.
  • The Navier-Stokes / Millennium Prize anecdote is not adjudicated here. Treat it as Altman's claim pending separate Record or Ground Truth work.

What policy teams should archive

  1. Policy and public-affairs teams: archive the OpenAI remarks URL, the UN News briefing page, and the UN Web TV 10228th-meeting asset. Quote the standards and incident-reporting asks when stakeholders ask what OpenAI told the Council.
  2. Safety / eval leads: map Altman's four international asks (capability measurement, risk assessment, safeguard sufficiency, human oversight) onto your existing eval plan. Flag gaps where you lack shareable evidence formats.
  3. Do not brief "the UN regulated RSI today." The documents here are remarks plus a news briefing. They are not adopted Council language.
  4. If you need multi-lab balance, wait for primary statements from other briefers before writing a second desk piece.