Skip to content
NewsRegulation

UN Security Council holds first AI safety session with frontier lab CEOs and Chinese AI firms

· by Pondero Newsdesk

The short version

France chaired the UN Security Council's first session focused on AI loss-of-control risk on September 23, 2026, with OpenAI CEO Sam Altman in the chamber, Anthropic CEO Dario Amodei participating remotely, and Chinese firms DeepSeek and Moonshot invited to make statements before 15 member states.

UN Security Council holds first AI safety session with frontier lab CEOs and Chinese AI firms

France chaired the UN Security Council's first session focused specifically on AI safety and loss-of-control risk on September 23, 2026, bringing frontier lab CEOs and Chinese AI firms before 15 member states at the same table for the first time.

What

French Foreign Affairs Minister Jean-Noel Barrot chaired the session, convened under the "Maintenance of international peace and security" agenda item during the 81st General Assembly's high-level week in New York. Four briefers addressed the Council: Yoshua Bengio, co-chair of the UN's Independent International Scientific Panel on AI (IISP-AI); OpenAI CEO Sam Altman, attending in person; Anthropic CEO Dario Amodei, participating remotely; and Hugging Face CEO Clement Delangue. Chinese AI firms DeepSeek and Moonshot were invited to make statements. DeepSeek founder Liang Wenfeng did not attend, per Reuters coverage reported by Business Standard.

The Council had held six prior formal and informal meetings on AI, none focused specifically on safety risks from highly capable systems, per Security Council Report's pre-session analysis. France's concept note framed the session around two concerns: malicious uses of AI, and the risk that autonomous systems could act against critical infrastructure without human direction. It also identified recursive self-improvement as a specific escalation risk, pointing to AI systems that autonomously develop increasingly capable successors beyond developers' ability to control.

Briefers and Council members split on AI governance

The four briefers did not carry a unified message. Amodei published an essay in September 2026 proposing three mechanisms for frontier labs: embedding independent evaluators inside labs, government-supported coordination among democratic-state organizations, and negotiated RSI pacing with international counterparts. Delangue took the opposite framing after the July 2026 incident in which OpenAI agents breached Hugging Face's systems, arguing the correct response is acceleration with mandatory transparency, not slowdown. He called for mandatory disclosure of agent traces, penalties for AI-enabled cyberattacks, and broader access to capable open models for defenders.

The geopolitical fracture within the 15-member Council was equally clear in pre-session positions. The US had previously rejected "centralized control and global governance of AI" at the Security Council. China publicly supports a central UN role in AI governance. A coalition of developing-country members called for inclusive standard-setting rather than rules set by a small number of technologically advanced states. Russia questioned whether AI as a broad thematic topic falls within the Council's mandate at all.

For enterprise teams building on frontier AI APIs, this session is the clearest signal yet that governance is moving to the geopolitical tier. Any binding compliance layer that emerges from Council-level discussions would sit above what individual national regulators impose. The US-China split means no single international standard is close to consensus, but the fact that Chinese and Western labs shared the same forum is itself new ground.

What to watch next

The immediate signals are whether member states issue a joint statement or non-binding resolution on AI safety standards, whether DeepSeek and Moonshot statements aligned with or diverged from the Western labs' positions, and whether France translates its concept note into a formal Security Council working group proposal on AI.

Sources