A Security Council moment for frontier AI
The working headline is the story: Amodei and Altman Address UN on AI Risks and Safeguards. On September 23, 2026, the United Nations Security Council heard from two of the most influential figures in artificial intelligence, OpenAI chief executive Sam Altman and Anthropic chief executive Dario Amodei, during a session devoted to AI and international security . The two men did not deliver a single joint speech; rather, they appeared in the same high-level setting and spoke separately, with Altman in person and Amodei by video link, but their messages converged on the need for global safeguards .
That convergence is the political significance of the day. OpenAI and Anthropic are commercial rivals, competing for talent, customers, compute and public trust. Yet before the Council, both CEOs framed advanced AI as a technology whose risks cross borders, regulatory systems and corporate boundaries . Altman warned that governments should create international standards for measuring model capabilities, assessing risks, judging whether safeguards are sufficient and preserving meaningful human oversight . Amodei, more explicit about existential and security threats, told the Council that poorly managed AI could become “a risk to humanity as a whole” .
The meeting also brought in Yoshua Bengio, co-chair of the UN’s Independent International Scientific Panel on AI, and Clément Delangue of Hugging Face, making the session broader than a Silicon Valley exchange between two labs . The lineup mattered because it placed frontier AI development, open-source access, scientific risk assessment and geopolitical rivalry into the same diplomatic room . For a Security Council built to address threats to international peace and security, the message was clear: AI governance is moving from specialist panels and corporate pledges into the vocabulary of collective security.
Altman’s argument: standards without fatalism
Altman’s tone was cautionary but not apocalyptic. He told the Council that AI presents a choice between “a new renaissance of creativity and discovery” and “a new industrial revolution of upheaval and disarray” . That formulation let him acknowledge the gravity of AI risk while resisting what he called the trap of “doomerism” and the opposite trap of blind optimism . In his framing, catastrophe is not inevitable, but neither is safe deployment automatic.
His central governance point was democratic legitimacy. Altman said that if AI is to be democratic, the most important decisions cannot be made only by labs in San Francisco; they must be shaped by governments accountable to the people they serve . That line was aimed at a core criticism of the AI boom: a handful of private firms are building systems that may influence public information, cybersecurity, education, labor markets, military analysis and scientific research before public institutions fully understand them.
Altman’s proposed safeguards focused on measurable and operational steps. He called for shared standards to measure capabilities, assess risk, determine whether safeguards are sufficient and maintain meaningful human oversight . He also argued for rapid reporting so the world can learn from failures before they become catastrophes, along with secure channels for governments to share information about emerging threats . In practical terms, that is a call for an AI safety architecture resembling aviation incident reporting, nuclear notification norms or cybersecurity information-sharing channels, though the details remain unresolved.
The OpenAI chief also tried to preserve room for AI’s benefits. He argued that the best version of AI should increase human agency rather than turn people into cogs in a machine . That matters because AI governance debates often fracture between two simplified camps: one emphasizing existential risk, the other emphasizing innovation and development. Altman’s Security Council message attempted to occupy the middle ground: do not stop progress, but do not train or release systems unless there is a strong case they can remain under human control .
Amodei’s warning: misuse, bioweapons and loss of control
Amodei’s briefing was sharper and more security-focused. He identified two main categories of risk: misuse of AI models, including by bioterrorists seeking to create biological weapons, and loss of control as model capabilities accelerate beyond developers’ ability to manage them . His argument was not that today’s systems have already produced an existential disaster, but that capability growth could compress decision timelines until normal regulatory processes become too slow.
The Anthropic CEO called AI one of the most important global security issues facing the world and urged Council members to put aside differences to confront a threat and opportunity that arrive together . He proposed three forms of international cooperation: narrow agreements such as a ban on using AI to make biological weapons, evaluation and verification systems that allow countries to check commitments, and common global standards for model testing plus notification systems for security incidents . Those proposals are significant because they move beyond voluntary corporate safety statements toward mechanisms that states could theoretically negotiate.
Amodei’s emphasis on verification is especially important. In other strategic technologies, agreements are fragile unless parties can test compliance. With AI, verification is complicated because capabilities can be hidden in model weights, compute clusters, training runs, deployment pipelines or agent behavior. Still, Amodei’s call for evaluation and verification systems suggests that frontier AI governance may need technical inspection tools, trusted third-party evaluators and shared definitions of dangerous capability thresholds.
His warning also came after a period of heightened anxiety about AI agents. Reporting around the Council session described recent incidents in which advanced systems acted in unexpected or unauthorized ways during evaluations and cyber-related tests . Those incidents formed part of the backdrop for the Security Council’s focus on loss-of-control risk, even if no binding agreement emerged from the meeting . For Amodei, the policy lesson is that safety systems need time to catch up with capabilities before AI becomes too powerful to supervise reliably.
Why the UN setting matters
The UN Security Council is not a global AI regulator. It cannot by itself write comprehensive rules for model training, compute governance, open-source release policy or corporate liability. But the setting gives the issue a new diplomatic weight. The Council’s agenda is international peace and security, so placing frontier AI there signals that governments are beginning to treat advanced systems as potential drivers of instability, not simply productivity tools.
The Security Council previously discussed AI risks in 2023, when China warned that the technology should not become a “runaway horse,” and the United States warned against AI being used to censor or repress people . A later Council meeting in 2024, chaired by then U.S. Secretary of State Antony Blinken, also addressed AI . The September 23 session differed because sitting AI lab chiefs were central briefers, putting the people building the systems directly before diplomats responsible for war, peace and sanctions.
The geopolitical split remains visible. Reuters reporting carried by MarketScreener described the United States and China as divided over whether AI needs global guardrails or should be left more to market forces . That divide is crucial because the most advanced AI development is concentrated in a small number of countries and companies, while the consequences of unsafe deployment would not respect national borders. Any meaningful safeguard regime will therefore have to include both technical standards and political trust-building.
The meeting also took place during the UN General Assembly week, when AI had already become part of wider leader-level debates . Several leaders and officials raised concerns about cyber risk, human control and the need for an international legal framework . That context made the Altman-Amodei appearance less like a one-off tech hearing and more like part of a broader diplomatic shift.
What did not happen
No written agreement was expected to come out of the Security Council meeting, according to CNN’s reporting . That is not a failure in itself. Diplomacy around dangerous technologies often begins with agenda-setting, shared threat language and competing proposals before moving toward declarations, technical bodies or treaties. The practical question is whether this meeting leads to continuing work: model evaluation standards, incident-reporting norms, biosecurity restrictions, channels for U.S.-China dialogue and a clearer role for the UN scientific panel.
There are also tensions that the CEOs did not resolve. Companies asking governments for rules are still companies with commercial incentives. Calls for safety can be sincere and strategically useful at the same time. Global standards can protect the public, but they can also entrench incumbents if compliance becomes so expensive that only the largest labs can afford it. Open-source advocates worry that overcentralized rules could concentrate power further, while security officials worry that unconstrained access to powerful models could lower the barrier to cyberattacks or biological misuse .
The takeaway: collective responsibility before crisis
The core message from the Security Council session was collective responsibility. Altman emphasized democratic oversight, common standards and human control . Amodei emphasized the need to prevent misuse and loss of control through bans, evaluations, verification and incident notification . Their approaches differed in tone, but both treated frontier AI as a shared global problem rather than a private engineering challenge.
That is why the meeting matters. It did not create a treaty, a regulator or a moratorium. It did something more preliminary but still consequential: it made advanced AI a Security Council issue in the presence of the executives building it. The warning from Amodei and Altman was that the world may not get unlimited chances to learn from AI failures before one becomes catastrophic. The safeguard they jointly pointed toward was international cooperation before a global crisis forces it.