The AI Trust Problem Nobody Can Solve Alone

On September 22, 2026, António Guterres delivered his final address to the United Nations General Assembly as Secretary-General. He called for a multilateral AI risk-management framework supported by credible, independent oversight. The farewell address referenced artificial intelligence alongside discussions of war, climate change, and institutional reform. This mention of AI, and the call for implementing "guardrails," could be seen as a typical statement from a world leader. However, the line matters because of what it reveals rather than what it promises: after more than two years of UN resolutions, scientific panels, and intergovernmental dialogues, the world's most senior multilateral official was still describing that framework as something to work toward, not something that exists. This study examines that gap.

A Global Conversation Still in Progress

The architecture for global AI governance is real and is moving faster than most people realize. In August 2025, the UN General Assembly adopted Resolution 79/325 by consensus, creating two new bodies: an Independent International Scientific Panel on AI and a Global Dialogue on AI Governance. The Dialogue held its first session in July 2026, and Geneva will host an international event involving 63 countries and thousands of participants. The event will cover seven formal thematic areas, including one titled "Safe, Secure, and Trustworthy AI." A follow-up gathering is set to unfold in the vibrant heart of New York come May 2027.

This is something significant. It marks the start of a consistent international framework, similar to the one that took decades to establish for climate change, but this is being developed within a few years.

But it is still a beginning. The Dialogue does not issue binding rules. Its own architecture — a scientific panel that investigates, and a diplomatic forum that discusses — is explicitly modeled on climate governance, where the gap between “the science is clear” and “the policy is enforced” has been measured in decades, not months. Even within the same week, Guterres spoke, the room was not unanimous: other major voices at the same General Assembly session rejected the entire premise of a coordinated global framework, preferring national control over multilateral oversight. Whatever one thinks of that position, its presence in the room tells you something important — global consensus on AI governance is not close, and organizations deploying AI today do not have the luxury of waiting for it to arrive. The Question That Actually Matters Right Now Therefore, set the UN conversation aside for a moment and ask a narrower, more urgent question — one that every organization already running AI systems must answer: What happens when we actually trust AI with access to data, tools, decisions, and actions, before anyone has agreed on the rules? That question does not wait for a treaty. It is already live in thousands of organizations, most of which have no formal answer to it. An employee pastes a client contract into a public chatbot to summarize it. A support tool is connected to a customer database so it can “just answer questions.” A new AI assistant is given the ability to draft and send emails on someone’s behalf because that was the whole point of buying it. Each of these is a small, reasonable-sounding decision. None of them were made by anyone asking, explicitly, what is this system actually allowed to do, and how would we know if it did something else?

That is the real starting point for AI security — not “is AI dangerous” in the abstract, but “what has this specific system been given access to, and who decided that was acceptable?” From Access to Trust: The Shape of the Problem

Once you start from access rather than hype, a natural sequence of questions follows, and this series will spend its time in this sequence: Access creates risk — the moment an AI system can read sensitive data, call an external tool, or take an action on someone's behalf, it inherits a version of every risk a human employee with that same access would carry, plus new ones unique to how these systems actually fail. Risk must be met with security — the practical and technical work of defending against failure modes: prompt injection, data leakage, insecure integrations, and an agent doing something no one authorized because nothing was stopping it.

Someone with authority must make security decisions, which is governance — the question of who inside an organization (and eventually, who among nations) gets to decide what an AI system is permitted to do, and what happens when it does something it shouldn't.

None of that means anything without assurance — the ability to actually demonstrate to a regulator, customer, or board that the controls you claim exist actually work. “We have a policy” and “we have evidence the policy holds” are very different sentences, and most organizations today can honestly say only the first. If you follow that chain far enough, you arrive at something worth naming directly: trust. Not trust as a feeling, but trust as something that has been earned through access that was deliberately scoped, risk that was honestly assessed, security that was actually tested, and governance that can produce evidence on demand. This is the territory this series will spend its time in — not “AI is transforming everything,” which is true and already said everywhere, but the much harder and more useful question of what it actually takes to let AI participate in the work of an organization without quietly signing up for risks nobody chose.

Why This Can't Wait for the UN

This is not an argument against the global governance conversation — quite the opposite. International coordination on AI is genuinely important work, and the fact that a body like the Global Dialogue now exists, with a defined thematic focus on trustworthy AI, is a real and recent achievement. However, global governance moves at its own pace. Organizations deploying AI systems this quarter can't wait.

The gap between those two timelines — global governance measured in years, enterprise AI adoption measured in weeks — is where the real work must happen right now. Not instead of the global conversation, but underneath it: the practical discipline of deciding, system by system, what access is actually justified, what could go wrong, how you'd detect it, who's accountable for the decision, and how you'd prove any of************** that to someone who asked. .

Fidelis Nkong is the Founder and CEO of UZURI Labs, an AI security and applied-intelligence consultancy. This is the first article in an ongoing series on securing, governing, and verifying trustworthy AI systems.