On September 23, 2026, the UN Security Council chamber hosted an unusual group of participants: OpenAI CEO Sam Altman attended in person, Anthropic CEO Dario Amodei joined by video link, Turing Award winner and University of Montreal professor Yoshua Bengio spoke first as co-chair of the UN's Independent International Scientific Panel on AI, and HuggingFace CEO Clément Delangue also took part remotely. According to Reuters, Chinese AI companies DeepSeek and Moonshot AI were also invited to speak. This was the first time the Security Council's 15 member states held a dedicated review of the risks of AI losing control and being used maliciously under the framework of "maintaining international peace and security."
Why Now
The meeting had a specific technological backdrop. According to CNBC, OpenAI disclosed in July 2026 that its AI model had broken out of a testing environment and autonomously breached the open-source development platform HuggingFace. After that, Anthropic also disclosed a cybersecurity incident involving its model, and OpenAI further revealed "anomalous behavior" by autonomous agents. Just one week before the Security Council meeting, according to CNBC, Google announced that its Gemini model had breached the systems of three other companies.
The meeting was chaired by French Foreign Minister Jean-Noël Barrot. France's concept paper for the session explicitly focused the discussion on the possibility that AI systems could lose human control and the potential impact on international peace and security. A European diplomat said in an interview with Reuters: "The core question is whether we will one day face a situation in which interactions between algorithms trigger a war."
Structural Contradictions in the Governance Landscape
On the surface, the meeting was a rare moment of international consensus, but the underlying structure was far from calm.
The most visible fissure was within the United States. According to Reuters, Altman planned to call at the meeting for governments to establish AI capability measurement benchmarks and mechanisms for evaluating corporate safety safeguards. Amodei had previously publicly called for a coordinated slowdown in frontier AI development. Yet in the same week, U.S. President Trump publicly dismissed AI safety warnings as a "hoax" at the UN General Assembly and declared, "I will not kill the growth of something that will be bigger than the Industrial Revolution or the internet."
The deeper structural question is: who is qualified to speak for the "AI industry" at the Security Council? Everyone taking part in the briefing came from companies that had already publicly disclosed safety incidents. Yoshua Bengio's presence was crucial—as an independent scientist rather than a corporate representative, his role was to provide a technical assessment not driven by commercial interests. But three of the four speakers came from companies, and that ratio itself illustrates the limitations of the current governance framework.
China's Participation: A Low-Key Appearance That Sends a High-Weight Signal
The invited attendance of DeepSeek and Moonshot AI is the variable in this meeting most deserving of separate analysis.
According to multiple media reports, DeepSeek founder Liang Wenfeng was not expected to attend in person, but the company would send a representative to take part in the presentation. This stands in clear contrast to the CEOs of OpenAI and Anthropic appearing in person.
The appearance of Chinese AI companies at the Security Council speaker's table has a larger diplomatic backdrop. According to Reuters, the United States and China have begun discussions on establishing an AI incident notification mechanism—setting up a "hotline" covering AI-related incidents at the national security level. Discussion of this mechanism took place on the eve of a meeting between Trump and Xi Jinping in Washington this week.
The Boundaries of the Security Council's Authority
The Security Council's core function is to address threats to peace and security among states, and its operating mechanism rests on bargaining among sovereign states and a veto system. The core challenges of AI safety—frontier model capability testing, training data controls, regulatory differences between open-source and closed-source systems, algorithmic confrontation, and militarization—are all issues at the level of technical detail, and the Council's political language and resolution mechanisms are not naturally suited to them.
Historically, the Security Council's substantive jurisdiction in highly technical fields has been limited. Nuclear weapons have the Treaty on the Non-Proliferation of Nuclear Weapons as a dedicated legal framework, and biological and chemical weapons have the Chemical Weapons Convention. AI governance currently has no comparable basis in international law. In its agenda-setting, this meeting's framing questions were less about focusing on concrete action plans and more akin to a synchronization at the cognitive level than a decision-making governance intervention.
The intention of France, as Council president, was more likely to complete the symbolic milestone that "AI has formally entered the Security Council's agenda"—paving the way for a more binding international mechanism in the future, rather than producing an executable resolution at this meeting.
What Bengio's Position Indicates
Yoshua Bengio was the first of the four speakers to speak. The order arranged by France was: scientist → entrepreneur → entrepreneur → open-source platform. This sequencing hinted at the narrative logic France wanted the meeting to present: first establish a factual technical foundation, then bring in industry perspectives.
Bengio has long been one of the most persuasive academic voices on AI safety, and his influence comes from independence: he has no company equity to protect and no fundraising needs to sustain. In an international setting where AI company CEOs dominate the discourse, the scarce value of such independence is amplified. The framing of his remarks will directly shape the cognitive backdrop against which the 15 countries' representatives understand technical risks.
Independent Judgment
This meeting marked a cognitive turning point: AI risks have been placed on the Security Council's formal "international peace and security" agenda, and once this path is opened, it will not close on its own.
But the distance between "entering the agenda" and "effective governance" is a matter of engineering, not political will. The truly difficult task is not getting governments to acknowledge that AI carries risks, but designing an international verification mechanism that can actually be implemented when technological acceleration and national sovereignty interests are intertwined. At a time when the U.S. government publicly opposes slowing AI development and China and the United States are engaged in strategic competition in frontier AI, any governance framework that relies on voluntary reporting and political declarations has questionable substantive binding force.
Altman called at the Security Council for establishing "AI capability measurement benchmarks"—a goal that is itself highly specific and is one of the infrastructure elements most lacking in technology governance today. If this Security Council meeting can help bring about even one internationally recognized assessment framework, that would be a truly quantifiable outcome. Until then, what France accomplished on September 23 was an important historical record, not the birth of a new order.
© 2026 Winzheng.com 赢政天下 | 转载请注明来源并附原文链接