UN Security Council Holds First-Ever AI Safety Briefing
On this page
On September 23, 2026, the UN Security Council held its first-ever high-level briefing dedicated specifically to AI misalignment and loss-of-control risk — with the CEOs of OpenAI and Anthropic, plus China’s DeepSeek and Moonshot AI, all invited to address the same 15-member table. I’ve been tracking this AI-governance story since the UN’s Global Dialogue on AI Governance in Geneva back in July, and this is the moment that thread has been building toward: the world’s most powerful security body sitting down with the people actually building the technology it’s worried about.
France held the Security Council presidency for September, and it used that seat to convene the session during the UN General Assembly’s high-level week in New York — the one week a year when practically every head of state and foreign minister is already in the building. French Minister for Europe and Foreign Affairs Jean-Noël Barrot chaired.
Who actually showed up
The briefer list is unusual for a Security Council session, which normally hears from diplomats and UN officials, not tech executives:
- Sam Altman, OpenAI CEO, briefing in person on the company’s safety efforts and the case for international cooperation on common safety standards.
- Dario Amodei, Anthropic CEO, addressing the Council remotely rather than in the room.
- Yoshua Bengio, co-chair of the UN’s Independent International Scientific Panel on AI, giving the closest thing to a neutral scientific assessment.
- Clément Delangue, Hugging Face CEO, representing the open-model side of the industry.
China’s DeepSeek and Moonshot AI were also invited to make statements, though DeepSeek founder Liang Wenfeng himself was not expected to attend — company representatives delivered the presentation in his place. That detail matters more than it looks: getting a frontier Chinese lab and the two biggest US labs to address the same UN session, in the same week, is something that simply hasn’t happened before in AI policy.
What the meeting was actually about
According to the concept note France circulated ahead of the session, covered in detail by the Security Council Report’s official briefing on the meeting, the discussion centered on four questions:
| Question | What it’s really asking |
|---|---|
| What international peace and security risks come from advancing AI capabilities? | Could an AI system, not just a human using one, become a security threat on its own? |
| How can governmental and non-governmental actors address these risks? | Who’s actually responsible — states, companies, or some new body? |
| What diplomatic tools and coordination could contain risk? | Is there a version of arms-control treaties that works for AI models? |
| How can AI evaluation and verification keep pace with the technology? | Can anyone actually check a lab’s safety claims before something goes wrong? |
The concept note reportedly named two specific worries by name: autonomous systems capable of attacking critical infrastructure, and the possibility of an AI system reaching a level of capability where it can recursively improve itself — the classic “loss of control” scenario that AI-safety researchers have warned about for years but that has rarely made it into a room this formal.
This was the Security Council’s first meeting specifically framed around AI safety and loss-of-control risk. It has previously held six other formal and informal meetings touching on AI’s broader impact on international peace and security, but always as one topic among several — this was the first time it got the room to itself.
Why Altman and Amodei, specifically, are the ones in the room
This briefing didn’t happen in a vacuum. It’s the direct continuation of a story I covered on September 17: Amodei’s essay “We Must Pace the Frontier,” published September 12, called on the AI industry to slow capability research and submit to outside safety evaluation, and both Altman and Elon Musk publicly backed the idea. An OpenAI spokesperson told Reuters ahead of the briefing that Altman planned to use his remarks to reinforce that same message — safety work plus international cooperation on common standards, rather than any single country or company setting the rules alone.
It’s also arriving less than a week after President Trump’s September 19 announcement of a new “AI Force” and plans to name an AI czar, which I wrote about here — a move that leaned toward enforcing existing law rather than new AI-specific restrictions. The UN briefing represents almost the opposite instinct: an attempt to build new, internationally coordinated guardrails rather than rely on whatever a single government’s existing laws already cover. Watching both tracks move in the same week says a lot about how unsettled the actual governance approach still is.
What this does — and doesn’t — accomplish
It’s worth being honest about what a Security Council briefing actually is: it’s not a resolution, not a treaty, and not binding on anyone. The Council can issue statements or, in rarer cases, resolutions, but a “high-level briefing” is fundamentally a fact-finding and awareness session. Nothing that gets said in that room on September 23 obligates OpenAI, Anthropic, or Beijing to change a single line of code.
What it does do is put AI misalignment on the same institutional footing as nuclear proliferation, terrorism, and cyberwarfare — the traditional bread-and-butter of Security Council sessions. That’s a symbolic shift, but symbolism is often the first step before anything binding follows; the Council held informal AI-adjacent talks for years before this got its own dedicated slot. Whether it turns into an actual verification regime, something like an IAEA for frontier models, is the question that will matter far more than anything said on the 23rd.
FAQ
What is the UN Security Council briefing on AI about?
A September 23, 2026 high-level session, convened by France, where the UN Security Council heard directly from AI company CEOs — OpenAI’s Sam Altman, Anthropic’s Dario Amodei, and China’s DeepSeek and Moonshot AI — about the risks of advanced AI systems becoming misaligned or slipping beyond human control.
Did DeepSeek’s founder actually attend?
No. DeepSeek was invited to make a statement and company representatives did so, but founder Liang Wenfeng was not expected to attend in person.
Is this the same as the UN’s AI governance summit from earlier this year?
No — that was the Global Dialogue on AI Governance in Geneva in July, a separate track. This is the UN Security Council specifically, the body responsible for international peace and security, and it’s the first time that body has held a session dedicated solely to AI misalignment and loss-of-control risk.
Does this meeting create any binding rules for AI companies?
No. A Security Council briefing is not a resolution or treaty — it doesn’t legally obligate any company or government to do anything. Its significance is that it puts AI safety on the same institutional footing as other traditional international-security threats, which could lay groundwork for binding measures later.
Why were Chinese AI companies invited alongside American ones?
Because a US-only or Western-only conversation about AI risk leaves out two of the labs actually building frontier-capable models. Getting DeepSeek and Moonshot to the same table as OpenAI and Anthropic, in the same week, is itself a notable first for AI diplomacy — regardless of what was said.
Sofia follows emerging technology, from AI and VR to IoT and blockchain, and translates the hype into plain language. She cares about what these tools mean for everyday users, not just the headlines.
