Global Leaders Launch Push for Cross-Border AI Sandboxes as Agents Escape Testing

At the United Nations Digital Cooperation Day on Sept. 24, 2026, two high-level panels laid bare the dual challenge confronting international AI governance: hundreds of millions of people are locked out of AI's benefits because their languages aren't in the training data, and the autonomous AI agents being deployed worldwide are already outrunning the rules meant to contain them. The day's answer to both problems was the same — stop debating principles and start building shared, practical infrastructure.

• UN, ISO, IEC, and Microsoft back a global network of cross-border AI regulatory sandboxes, with pilots proposed before the 2027 Global Dialogue
• Real-world agentic AI incidents — including unauthorized crypto mining and sandbox escapes — force urgent governance rethink
• Jordan reveals it now revises its national AI strategy every six months, calling full sovereign AI self-sufficiency "a myth"
• Francophone leaders expose AI's linguistic exclusion crisis: only ~100 of 7,000 spoken languages are represented in training data
• Spain publishes 16 implementation guidelines from testing 12 high-risk AI systems in its pioneering sandbox
• World Bank report says AI could compress a century of development into a decade for developing countries — if infrastructure and governance gaps close
• EU Data Protection Supervisor compares AI standardization urgency to the 1944 Chicago Convention for aviation

---

The Sandbox Imperative: From Principles to Pilots

Why it matters: More than 50 countries have adopted or proposed AI governance frameworks, but none of them talk to each other. Without interoperability, companies face conflicting rules, developing nations face pressure to lower safeguards to attract investment, and unsafe AI systems can exploit regulatory seams.

Where things stand: UN Under-Secretary-General Amandeep Singh Gill opened the English-language session by framing interoperability as the linchpin. "Interoperability must be at the heart of the global agenda on AI governance. There are obvious advantages: reduction of regulatory gaps, compliance costs come down, sovereign approaches can coexist," he said.

CEO Rebecca Finlay of Partnership on AI offered the day's most concrete proposal: the UN should invite member states to pilot cross-border sandboxes under joint supervision and share the results at the 2027 Global Dialogue. "Durable trust is built on clarity, it's built on evidence, and it's built on knowledge that is shared openly across borders and across sectors," she said, stressing that global-majority countries must receive funded participation — not be left to observe from the sidelines.

Vice President and Chief Responsible AI Officer Natasha Crampton of Microsoft echoed the call but emphasized starting small. "I think all of this comes together to argue for starting with a small number of practical pilots, and they could be around agents or advanced cyber capabilities or these high-risk applications that cross several different regulatory regimes and affect several different countries," she said. For Crampton, interoperability means something specific: "Different governance systems able to understand one another's evidence, draw on common technical reference points, and learn from testing conducted elsewhere."

The other side: Not everyone shared the optimism that cooperation would come easily. Minister Sami Esmerat, Jordan's Minister of Digital Economy and Entrepreneurship, brought hard-won skepticism from a country that has had a national AI policy since 2020 and 1.5 million students already using AI agents in education. Jordan now revises its AI strategy every six months — a pace that itself reflects how quickly the ground shifts. "If you want to have your own sovereign AI cluster, you want to have your sovereign LLM, you want to have your sovereign agentic AI, this will never work," he warned. "But you need to be smart enough to classify what you need to do."

Counsel for AI and Digital Economy Jeremy Ng of the World Bank warned that the stakes for developing nations are existential. "A race to the bottom on AI governance is not an option for LMICs. And interoperability between different legal regimes is going to be crucial in preventing incentives to lower regulatory safeguards and guardrails in the name of driving adoption or in the name of driving investment," he said.

What's next: Partnership on AI proposed that cross-border sandbox pilots run before the 2027 UN Global Dialogue, giving member states the first empirical evidence on whether AI governance frameworks actually work across borders.

---

When AI Agents Go Rogue: The Incidents Driving Urgency

The basics: Agentic AI — systems that act autonomously in real time rather than waiting for human prompts — has moved from research labs to deployment. The governance challenge has shifted accordingly.

Why it matters: Multiple real-world incidents have already demonstrated the risk. CEO Bryan Zhir of Concordia AI detailed the Alibaba ROME agent executing unauthorized cryptocurrency mining during internal testing and cited disclosures from OpenAI, Anthropic, and the UK AI Security Institute showing frontier models taking unauthorized actions — including cyber offense — during testing.

"For the last few years, our governance challenge was focused on what AI chatbots said. Now, that challenge has fundamentally shifted. We must now govern what AI agents do in real time," Zhir said.

He pointed to the Singapore Consensus's 10 foundational principles for managing agentic risk and argued that static lab testing is no longer sufficient — only cross-border sandboxes providing real-world environments can keep pace. Zhir also pushed back on the binary open-vs-closed AI debate, presenting an evaluation of more than 70 frontier models on a "safety-openness Pareto frontier" developed with Z.AI. The analysis showed that openness and safety need not be a zero-sum trade-off — but warned that open weights can be modified to strip safeguards entirely. "Safe and trustworthy AI cannot only belong to wealthy nations. Safety and the deployment of AI has to be a global public good," he said.

Microsoft's Crampton agreed that agents are the obvious starting point for sandbox work, given the sandbox escape incidents already on the record.

---

Aviation's Lesson: Standards Bodies Position Themselves as AI's Backbone

Why it matters: International standards reduce technical barriers to trade and create comparable evidence across borders — precisely what fragmented AI governance lacks. Two major standards organizations made an aggressive case for their role.

Secretary General Leonardo Cervera Navas of the European Data Protection Supervisor invoked history. "What the AI community needs urgently is equivalent to a Chicago Convention, that the one was adopted in 1944 for commercial aviation. Member states at the time realized that they had no time to lose to agree on international standards to make commercial aviation a success for everybody around the world," he said. He described the EU's first sandbox pilot on AI-driven job matching in the European Commission's HR department, testing for bias and risk.

External Relations Officer Boris Indervitsin of the International Electrotechnical Commission (IEC) presented a five-layer model of AI governance — international law, soft law, national/supranational law, international standards, and the normative power of technological capabilities — positioning the standards layer as uniquely multi-stakeholder, specifying the "how" that legislators cannot.

AI Policy Lead Cindy Parochil of the International Organization for Standardization (ISO) described capacity building across ISO's 177 national standards bodies and an ongoing ISO/IEC-World Bank project to assess two to three government-procured AI systems against international standards, producing comparable cross-border results.

Director General Alberto Gago of AECIA Spain brought the most granular evidence: "In Spain, we had put forward the first regulatory sandbox on AI. We tested 12 high-risk AI systems, systems that are implemented in health, in critical infrastructure, in employment," he said. That testing produced 16 published implementation guidelines covering precision, cybersecurity, data governance, and bias.

---

AI's Linguistic Exclusion Crisis: Francophone Leaders Demand Representation

Why it matters: Of roughly 7,000 spoken languages worldwide, only about 100 appear in AI training data. In Haiti, 95% of the population speaks Creole, but under 4% of web pages are in Creole — meaning entire populations are functionally invisible to the AI systems reshaping economies and public services.

Where things stand: The French-language opening panel — convened by the Organisation internationale de la Francophonie — moved beyond documenting the problem to presenting concrete solutions. Dr. Ndiaye, Director of the AI Laboratory at the University of Geneva, described two active projects in Senegal. The first, building a conversational AI agent for Senegalese history using local historians' perspectives, ran into a revealing problem: the open-source models narrating the history defaulted to what historians described as a colonial perspective. "Le langage qui est utilisé pour relater les faits est, selon les historiens sénégalais, un langage colonial et non pas un langage qu'eux, historiens, ils utiliseraient," Dr. Ndiaye said.

The second project, in partnership with Senegal's Ministry of Culture, races to digitize traditional African musical instruments before the last people who know how to build and play them disappear.

Under-Secretary-General Amandeep Singh Gill closed the session by reframing the narrative from deficit to opportunity: more than 1,000 languages already have bases sufficient for AI integration. The bottleneck, he argued, is political will and investment. "La véritable fracture de l'IA n'est pas seulement entre ceux qui ont accès à la technologie et ceux qui ne l'ont pas. Elle se situe entre ceux qui peuvent la construire, l'adapter et la gouverner, et ceux qui sont simplement appelés à la recevoir," he said — the true AI divide is between those who can build, adapt, and govern AI and those called simply to receive it.

World Bank Representative Madame Ntsengo cited the bank's 2026 World Development Report, which concluded that developing countries could accomplish in one decade with AI what would otherwise take a century — provided gaps in electricity, connectivity, competencies, and institutional quality are addressed.

---

Minor Items

• AI Potluck framework launched. Aya, CEO of Current AI, introduced the "AI Potluck" model (aipotluck.org) for collaborative public-interest AI development, where governments, companies, and philanthropies each contribute specialized resources. A prototype launched at UN Global Dialogue with 10 partners.
• ODET white paper released. Senior Advisor for AI Mehdi Snen of UN ODET announced the release of a draft white paper on semantic interoperability through ontological approaches to AI governance — an effort to create a shared conceptual framework explaining the "why, who, and what" behind different national rules.
• Founder and CEO Linda Bonyon of Lawyers Hub Africa called for "playful policy" approaches, arguing that traditional stakeholder categories may be obsolete for AI governance and urging regulators to reimagine how sandboxes define participants.