OpenAI CEO Sam Altman Calls for Extreme Caution on AI at UN Security Council
OpenAI CEO Sam Altman stated at the UN Security Council that the current development of artificial intelligence "requires extreme caution," and the industry should not accept excessive technological risks due to the significant benefits and importance that cannot be slowed down.
This public briefing by the 15-member Security Council was hosted by France, coinciding with the high-level week of the General Assembly. Attendees included Anthropic CEO Dario Amodei (via video link) and Hugging Face co-founder Clément Delangue. Altman spoke on-site, stating that the company would take a "middle path": neither falling into blind optimism nor apocalyptic thinking. He also mentioned that if AI is to be democratic, the most important decisions cannot be made solely by laboratories in San Francisco, but must involve governments accountable to the people and international cooperation; historically, even countries that do not like each other have come together for a powerful new technology.
Before the meeting, OpenAI told reporters that Altman would advocate for using benchmarks to measure model capabilities, assess company safeguards, and position himself as a "pragmatic centrist": aiming for common safety standards while continuing development to avoid excessive concentration of capabilities in a few companies. Security Council member states include the US and China, both of which will speak. A European diplomat stated that core issues include whether algorithm interoperability could trigger war, which falls under the Security Council's jurisdiction. UN Secretary-General António Guterres has emphasized the inadequacy of national-level governance and the need for global coordination.
Outside the venue, the rhythm was inconsistent. US President Donald Trump rejected global uniform rules, calling it a "global conspiracy," and stated that the US should encourage AI rather than restrain it. This is the third public meeting of the Security Council on AI, following sessions in July 2023 and December 2024. In the weeks leading up to the meeting, the industry made intense statements regarding the risks of losing control and whether to slow down; Amodei stated at the meeting that poor management could even endanger all of humanity.
In recent years, Altman has pursued two lines: one advocating for the US to lead in setting standards and arguing that laboratories should not expand at maximum speed indefinitely, while continuing to train stronger models and seeking global distribution. OpenAI has publicly discussed capability assessments, mandatory incident reporting, and the inability to safely achieve complete alignment in recursive self-improvement. Presenting the same set of arguments at the Security Council shifts the audience from Congress and the G7 to a veto-holding security body.
In market mechanisms, this is a regulatory expectation event. Buyers are US laboratories seeking global access and mutual recognition of standards; the pressure on sellers falls on companies that follow open weight routes and are unwilling to accept external testing. Funding does not immediately relocate due to a single briefing, but the Security Council stage will elevate "safety compliance" as the price of entry for going abroad. Beneficiaries are closed-source laboratories with existing evaluation and government affairs backgrounds, while those under pressure include narratives that equate slowing down with surrender and national industries excluded from the standard-setting table.
Additionally, Yoshua Bengio, as co-chair of the UN Independent International Scientific Group, also appeared in the briefing arrangements. The meeting itself is not expected to pass binding resolutions.
Source: Public Information
ABAB AI Insight
Altman views the Security Council as a standard output venue, not a training halt. What OpenAI needs is for the US to lead and other countries to follow in defining assessment metrics, allowing cutting-edge models to continue selling while establishing thresholds for "responsible laboratories." The syntax of the middle path aligns with his framework in the G7 and company blog: it is acceptable to slow down maximum speed, but development rights cannot be handed over to a single international stalling body. Trump's outright rejection of global solutions effectively pushes Altman's multilateral rhetoric back to domestic legislation and bilateral coordination.
The capital path is standards first, products later. Whoever defines capability benchmarks and safety checklists can decide which models can enter government, banks, and cross-border clouds. Anthropic is openly calling for speed control, Hugging Face represents open weights, and the three on stage are pushing three business models into the same safety narrative. OpenAI's government affairs line has shifted to support federal mandatory rules, independent assessments, and incident reporting for the strongest models, exempting open-source systems—this piles regulatory costs onto closed frontiers while retaining an open side in the ecosystem.
Similar structures appear in civil nuclear power, civil aviation airworthiness, and chip export controls: technology companies first approach international organizations for "common standards," then turn those standards into market access. The industry is at a stage of accelerating capabilities, unresolved alignment, and mutual distrust among countries. The third discussion of AI by the Security Council indicates that the topic has moved from an ethical forum to the international security agenda, but the absence of a resolution still shows that the power to define remains in the capitals rather than in New York.
Essentially, it is a transfer of pricing power amid regulatory changes. The mechanism is: when models are framed as capable of self-improvement and potentially threatening international security, assessments and incident reports become tolls; laboratories exchange "extreme caution" for political permission to continue training, while governments use the Security Council to gain inquiry rights over laboratories. The real brakes are not in the speeches, but in who has the authority to halt the next training and who retains market share after a halt.
ABAB News · Cognitive Laws
The more it is said that risks cannot be taken for profit, the more it indicates that profits are determining the pace.
The middle path is to keep running, just handing the ruler to standards that one participates in formulating.
The Security Council provides a stage, not a switch.