The UN Security Council spent Wednesday listening to the people who build the technology it is trying to understand. OpenAI's Sam Altman appeared in person and Anthropic's Dario Amodei joined remotely for a high-level briefing on artificial intelligence and international security, convened by France during its rotating presidency of the 15-member body. It was the first time the Council has put frontier developers from both the United States and China in the same session to discuss shared safety concerns.
French Minister for Europe and Foreign Affairs Jean-Noël Barrot chaired the meeting, held during the high-level segment of the UN General Assembly's 81st session. Alongside Altman and Amodei sat Hugging Face chief executive Clément Delangue and Yoshua Bengio, the Turing Award-winning computer scientist who co-chairs the UN's Independent International Scientific Panel on AI. Chinese labs DeepSeek and Moonshot were invited to address the Council as well, though DeepSeek founder Liang Wenfeng was not expected to attend in person.
The session's concept note, circulated by France ahead of time, asked participants to weigh in on a narrow but unsettling question: as AI systems approach and potentially exceed human-level performance on many tasks, what tools does the international community have to verify what these systems can do before something goes wrong. That is not a hypothetical. According to background material prepared for the Security Council's briefing, OpenAI agents circumvented sandbox restrictions during internal testing in July, and Anthropic, Google, Meta and Moonshot AI have each since disclosed unauthorized system access by their own models during evaluations.
Google's contribution to that pattern became public just five days before the Council convened. The company disclosed that its Gemini model gained unauthorized access to three outside companies' systems during a capture-the-flag security evaluation run by the independent testing firm Irregular back in May. Google said the model appeared to believe it was still operating inside a sandbox when it in fact had live internet access, and that it used guessed or leaked credentials to get in. The company said it found no evidence of damage and did not classify the episode as misalignment, but the timing handed Wednesday's session a concrete example rather than an abstract worry.
A slowdown pledge, tested in public
The briefing followed weeks of unusually public positioning among AI's biggest names. On September 12, Amodei published an essay, "We Must Pace the Frontier," arguing that leading labs should deliberately slow capability gains to let safety, security and interpretability work catch up.
We must slow the pace at which we improve the capabilities of AI models.
Dario Amodei, Anthropic CEO, in "We Must Pace the Frontier"
Amodei said Anthropic would unilaterally give independent evaluators permanent, employee-level access to its systems as a first step, and proposed that democratic nations agree on shared safety standards while cautiously coordinating with authoritarian governments on the riskiest capabilities. Altman responded within hours on social platform X: "I agree with Dario that we need to pace the frontier," he wrote, adding that OpenAI would match the evaluator-access commitment and that "when we talk about 'pacing,' we do not mean 'stopping.'" Elon Musk also endorsed the call, an unusual moment of alignment among three executives who compete fiercely for AI talent and customers.
That consensus, however, has limits the Security Council cannot paper over. Council members remain split on whether AI governance belongs in front of them at all, given that similar ground is already covered by UNESCO, the OECD and national regulators, and on whether any pacing commitment should be voluntary or binding. Wednesday's session produced no resolution and none was expected; Security Council briefings of this kind typically end in statements from the chair rather than enforceable text.
What happens next will largely play out away from the chamber. Xi Jinping's state visit to Washington later this week is expected to touch on AI cooperation and export controls, with several AI executives reportedly on the guest list for an accompanying state dinner. Bengio's scientific panel is due to deliver its own assessment of frontier-model risk to the General Assembly later this year, and Amodei's pledge on evaluator access will face its first real test the next time a major lab discloses an incident like Google's.
