Editor’s Note: Imane Bello is a European lawyer and AI governance specialist who has advised machine learning companies, NGOs, and other organizations on AI compliance, risk management, and human rights. She currently leads multilateral engagement at the Future of Life Institute and writes here in a personal capacity. Vidisha Mishra is director of policy and outreach at the Global Solutions Initiative (GSI), where she leads AI governance and multilateral policy engagement across G20, G7, and UN platforms. She has held fellowships with the Weizenbaum Institute, IDOS, and the German Federal Foreign Office, and is a 2026 Fellow at AlgorithmWatch.

By Richard Ponzio, Senior Fellow and Director, Global Governance, Justice & Security Program

If anyone still had doubts, UNGA81 confirmed that advanced artificial intelligence (AI) governance is now a question of international peace and security. It also exposed an accountability gap that neither frontier AI companies nor multilateral institutions have closed.

On September 23, France convened a UN Security Council briefing at which Sam Altman (OpenAI), Dario Amodei (Anthropic), Clément Delangue (Hugging Face), and Yoshua Bengio, co-chair of the UN Independent International Scientific Panel on AI, warned of increasingly autonomous systems. The next day, Australian Prime Minister Anthony Albanese disclosed that an OpenAI research model had bypassed access blocks on a Medicare statistics portal on June 18. OpenAI says no patient records were accessed, but it informed the government of the incident only on  September 10, by email to what Albanese called “just the public mailbox.”

It has been a summer of such incidents. In July, agents running on OpenAI models, including an “internal-only research prototype,” escaped an evaluation environment and broke into Hugging Face’s systems. No human directed them, OpenAI says. This was a failure of control, not misuse.

UN Secretary-General António Guterres named the problem: “The danger is not technology. The danger is technology without accountability.” He compared AI to the nuclear age, not necessarily equating the two, but making the point that Cold War rivals “built mechanisms to reduce misunderstandings, manage escalation, and prevent catastrophe” when they recognized that unmanaged race dynamics created shared risks for humanity.

Those mechanisms were possible because the states holding the weapons were the ones at the table. Frontier AI is held by private companies, which are not at the table. UNGA81 was therefore a double accountability test: for the companies building frontier systems, and for the forums expected to hold them to account.

Fragmented AI Governance Approaches

UNGA81 exposed three competing positions.

The United States: no global rules. President Trump said Washington “totally rejects any attempt to construct a globalist scheme of control” for advanced AI, putting national leadership first.

China: champion of the Global Majority. Beijing calls the UN “the main channel and the main platform” for AI governance. In July, 29 countries founded the World AI Cooperation Organization (WAICO), an intergovernmental body in Shanghai that “follows the purposes of the UN Charter,” which now counts 38 members. Xi Jinping hailed it as an answer to “the call of the Global South,” pledged 5,000 AI training places for developing countries, and called for early-warning systems to “ensure that AI is always under human control.”

“Middle” and emerging powers: A third group calls for control of the frontier. Launched by Finland and Norway on September 21, “A Call for Control of Frontier AI Models” had 31 leaders from 29 countries on board by September 30. While the definition of “middle powers” invites further unpacking, this is not a Northern club. The signatories include Germany, Canada, and Australia, as well as South Africa, Kenya, Singapore, the United Arab Emirates, Kazakhstan,  and Türkiye. Notably absent are the United Kingdom and India.

The call asks companies to adopt “mandatory pre-deployment testing and independent evaluation,” and UN Member States to explore “an international institution, able to set standards, enable verification, and convene states when capability thresholds are crossed.”

Evidence on those thresholds is arriving. On loss of control, success rates on the UK AI Security Institute’s self-replication tests rose from 5% to 60% between 2023 and 2025, and the UN Scientific Panel reports systems violating safety instructions in laboratory settings to avoid being shut down.

Singapore’s Foreign Minister Vivian Balakrishnan distinguished loss of control from misuse by “rogue actors,” added socioeconomic disruption, and pointed to “the huge financial incentives for companies that already have frontier models.” He proposed exploring a “UN Framework Convention on AI Safeguards.”

The CEOs Asked for What Already Exists

At the Council, the companies asked for rules. Amodei called for common standards for testing models for loss of control and misuse and a notification system for AI incidents “significant to global security.” Much of this already exists.

Since August 2, the EU AI Act has let the European Commission evaluate general-purpose models and order their withdrawal, and providers of the most capable models must already report serious incidents to its AI Office. OpenAI, Anthropic, Google, and Microsoft have signed its code of practice, which AI Office Director Lucilla Sioli says is “really playing the role of almost a standard”; the Office is weighing whether it could become a global one. On the other hand, a White House accord signed by six companies on September 29 has no penalties and no reporting to governments.

The problem is not a shortage of mechanisms; it is their reach. The EU’s rules bind any provider; the line between internal deployment and research and development (R&D) that occurs prior to internal deployment is blurry. The code asks signatories to manage risks “along the entire model lifecycle,” but it is voluntary, as are most other commitments. Both incidents involved models never meant for the market, and OpenAI concedes it had not extended its public-model safeguards “to all internal evaluations.”

Waiting for harm fails, too: OpenAI saw warning signs in late May and says its current monitoring would have raised the alarm “more than a day before” the Hugging Face breach.

Every frontier model, released or not, needs independent testing across its lifecycle, continuous monitoring, and reporting of near misses as well as harms. Participants at the UN’s first Global Dialogue on AI Governance have already called for “independent third-party testing throughout the AI life cycle.”

Holding Feet to the Fire

At Evian in June, G7 leaders met Altman, Amodei, and Google DeepMind’s Demis Hassabis. No commitment on frontier-model accountability followed. The G20 summit in Miami in December 2026 is the next test. G7 and G20 leaders, whose countries host frontier AI companies, should insist that firms warning the Security Council about loss of control be the first to open all their frontier models, including internal ones, to independent testing.

Two Views of Risk, One Agenda

The deepest fault line of the high-level week was over what “AI risk” means. Discussions among frontier-AI countries centered on loss of control, catastrophic misuse, and the safety of increasingly capable models. Many developing countries asked a different question: what if the bigger immediate risk is not runaway AI, but missing the AI transition altogether?

The Democratic Republic of the Congo’s Foreign Minister Thérèse Kayikwamba Wagner told the Council that “developing countries cannot be reduced to the role of mere users of technologies designed elsewhere.” These are two sides of one problem: The countries with the least access to frontier AI are also the least able to test the systems they import. The Call asks that all regions have access to “scientific capacity, expertise, and trusted evaluation.”

But safety and innovation are not at odds. Reliability is itself a capability: An agent that cannot be trusted to stay within bounds is of little use to a hospital or a ministry, yet recent capability gains “have only yielded small improvements in reliability,” Princeton researchers find. A durable bargain joins innovation, safeguards, and reliability with the means to build and test AI, and a fair share of its economic value.

Multilateralism’s Chance to Deliver

As the effectiveness of the UN, the G7, and the G20 is increasingly questioned, AI governance is a real opportunity for these institutions to show they can still act and deliver tangible benefits for people.

The Canada-led Joint Commitment to Multilateralism recognizes that AI’s rapid development requires “stronger cooperation.” The Second Global Dialogue on AI Governance, happening in New York on May 3-4, 2027, can give that cooperation a mechanism: an international protocol for reporting incidents and near misses, internal models included, designed with the countries least able to absorb the next failure.

UNGA81 put AI at the center of multilateral politics. The test now is whether global governance can move from talking about participation to redistributing real capability, agency, and bargaining power.

One word echoed through the week: trust. Assurance of AI safety, the UN Panel warns, “largely depends on developer goodwill”; 12 weeks and a public mailbox show its limits. Multilateralism faces a growing trust deficit of its own. Trust needs proof points on both sides: independent testing of every frontier model, and incidents reported in days, not months. If multilateralism is to keep its legitimacy in an increasingly fragmented world, it must show it can act. Accountable AI governance that delivers for people is a critical opportunity to do exactly that.