Altman Tells the UN Security Council That No Level of Catastrophic Risk Is Acceptable

OpenAI CEO Sam Altman addressed the United Nations Security Council on September 23, 2026, and published his remarks as delivered. The speech framed frontier AI as a choice between "a new Renaissance" and "a new Industrial Revolution of upheaval," named two specific failure modes the company says must be avoided, and asked governments to build a shared standards mechanism for measuring capabilities and verifying safeguards. It also contained a technical claim — that an OpenAI model solved one of the Millennium Prize Problems — presented without supporting detail.

The framing: agency, not automation

Altman argued that the timeline feels compressed, with recent model progress making both the upside and the risks more immediate. His central pitch was that AI should expand individual agency rather than absorb it: more ability to learn, create, build and participate, and a defense against the rising complexity of ordinary life. He said explicitly that there are things AI cannot and should not automate, and that human judgment and human values are on that list.

That is a rhetorical position, not a product commitment. The remarks did not specify which decisions OpenAI believes should remain off-limits to automation, how the company operationalizes that line in its own systems, or how it would be enforced across an industry that is aggressively shipping autonomous agents. For engineering and product teams, the gap between "people must remain at the center of AI decision-making" and a concrete design requirement is where most of the hard work lives, and the speech left it open.

Altman also acknowledged that some statements from people building AI have sounded dystopian enough to resemble bad science fiction — an unusual admission from a frontier lab CEO, and one aimed at the anxiety in the room.

Two failure modes, one of them about companies like his

Altman said he largely agrees with Yoshua Bengio, who also spoke, but named two ways things could go badly rather than three.

The first is loss of control: systems that move faster than people can follow or intervene. He tied this specifically to recursive self-improvement, arguing that as the process of building AI becomes more automated, progress could accelerate sharply. He said the industry must not accept excessive technological risk simply because the benefits feel too important to slow down, and that competitive pressure is not a justification for rash decisions. He stated that OpenAI has unilaterally slowed down in the past and will again — without giving an example, a date, or what "slowing down" meant in practice.

He was unusually blunt on risk thresholds, saying it does not matter whether one estimates the chance of catastrophe at 10 percent, 1 percent or 0.1 percent, because none of those levels is acceptable, and that models should not be trained unless there is an extremely strong case they can be kept under human control. That is a stringent standard on its face. It is also unfalsifiable as stated: the remarks did not define who judges whether the case is "extremely strong," what evidence qualifies, or what happens when an internal evaluation is contested.

The second failure mode is concentration of power. Altman said no person, company or country should be able to use the most powerful models to impose a worldview on everyone else, and that any actor claiming to be the only trustworthy steward of the technology can use that belief to justify almost anything. Coming from the CEO of one of the largest frontier labs, that is a claim worth holding onto — OpenAI itself is one of the small number of actors the argument implicates, and the speech did not address how the company's own structure, compute position or government relationships fit the critique.

The ask: complementary national and international standards

The concrete policy request was a mechanism for complementary national and international frontier AI standards covering four things: measuring capabilities, assessing risks, determining whether safeguards are sufficient, and preserving meaningful human oversight as systems become more autonomous. Altman said common standards would let countries compare evidence, verify compliance and share a common vocabulary.

He added two supporting asks: accurate and fast incident reporting with classification and reporting protocols, so failures can be learned from before they become catastrophes; and secure channels among governments, critical infrastructure operators and technical experts for sharing vulnerabilities and emerging threats.

He also set constraints on what such standards should not do — lock in incumbents or favor one business model — and said they must accommodate open and closed developers, new entrants and established labs. Each government, he said, should decide how to fold the standards into its own legal system.

That design is familiar from other technical governance regimes, and the hard parts are the ones the speech acknowledged least. Capability measurement is contested even among labs. Verification of compliance for closed-weight models typically requires some form of privileged access, and the remarks did not say what OpenAI would be willing to expose, to whom, or under what confidentiality terms. No venue, timeline, sponsoring body or funding source was identified.

Altman also said companies must not substitute for the democratic process, and that the most important decisions cannot be made by labs in San Francisco alone.

The Navier-Stokes claim

To illustrate capability progress, Altman laid out a four-step timeline: models that were okay at grade-school math three summers ago, pretty good at high-school competition math two summers ago, gold-medal level at the most prestigious international math competition last summer, and, a few weeks before the speech, a model that solved one of the Millennium Prize Problems — the Navier-Stokes equations. He noted those equations bear on aircraft design, weather forecasting and the study of blood flow.

This is the single most consequential factual assertion in the remarks, and the published text supplies nothing to evaluate it. No paper, preprint, proof artifact, reviewing mathematicians, model name or verification process was disclosed, and the speech did not say whether the result concerns existence and smoothness in the form the prize specifies or a related problem. Claims at this level are normally settled by lengthy peer review; readers should treat a podium statement as a claim awaiting that scrutiny, not as a confirmed result. For anyone modeling capability timelines, the difference matters enormously.

Why the venue matters

Addressing the Security Council rather than a standards body or a legislature puts frontier AI in the vocabulary of international peace and security. The practical effect is unclear: the Council's tools are geopolitical, and the speech asked for technical cooperation of the kind usually built through specialist institutions. No Council action, resolution or follow-on process was described in OpenAI's post.

What to watch

• Whether OpenAI publishes any technical documentation, proof artifact or third-party mathematical review supporting the Navier-Stokes claim, and how the mathematics community responds.
• Whether the company puts specifics behind "we have unilaterally slowed down" — which model, which decision, and what the trigger was.
• Whether any government or multilateral body picks up the standards proposal with a named venue and timeline, rather than a general endorsement.
• How the incident-reporting idea is scoped: what counts as a reportable AI incident, who receives reports, and whether disclosure is mandatory or voluntary.
• Whether OpenAI's stated opposition to power concentration translates into anything structural — access commitments, external audit rights, or governance changes — or remains a statement of principle.