AGI Won’t Arrive. It’s Already Being Assigned.

A year ago, writing about how artificial general intelligence would be distributed still required imagining the future. This month, it requires reading the news.

On April 7, 2026, Anthropic announced a model called Mythos and simultaneously announced that the public would not be allowed to use it. The reason given was that the system had proven capable of finding thousands of previously unknown security vulnerabilities across every major operating system and browser — including a twenty-seven-year-old flaw buried inside OpenBSD that a generation of engineers had failed to catch. Instead of a public release, Anthropic opened a private program called Project Glasswing, granting roughly forty organizations early access. JPMorgan, Goldman Sachs, Bank of America, Citigroup, Morgan Stanley. Amazon, Apple, Microsoft, Google. Nvidia, Broadcom, Cisco. You can read the list. You will not find your hospital on it, or your city government, or the open-source projects your bank actually runs on.

Mythos is not AGI. Nobody serious is claiming it is. But the shape of what just happened — a frontier capability developed in private, deemed too powerful for ordinary people, and then handed to forty pre-selected institutions by the unilateral decision of a single company in San Francisco — is the shape. It is the template. It is the answer, delivered quietly, to the question the AGI conversation has been avoiding for years.

Who gets it? Not you.

Most people still talk about AGI as if it were a kind of public event. One day, the story goes, we cross some invisible line, journalists announce it, CEOs tweet about it, and eventually everyone has a piece of godlike intelligence in their pocket, answering questions, solving problems, making life easier. It is a strangely comforting fantasy: the idea that when a mind greater than ours arrives, it will come as a consumer product. Something with a login screen and a dark mode and a slider for "creativity."

Stop and think about what true AGI would actually mean. A system that reasons across domains, plans over long horizons, designs weapons and drugs and software, models human behavior well enough to manipulate it at scale, and rewrites its own capabilities to improve them. You do not roll that out like a new messaging app. You lock it down. You fight over it. You classify it. And if you are being honest with yourself, you fear it.

If AGI arrives, it will not arrive evenly. It will emerge inside specific institutions, under specific power structures, with specific motives. Some people will live close to it. Most will only live downstream of it. The gap between those two groups may become the defining inequality of the century.

Every technology that has shifted the balance of power on Earth began in the hands of those who already had the resources to build it and the reasons to guard it.

Nuclear technology did not debut as clean energy for the masses. It debuted as weapons tested in deserts and dropped on cities. Over time, it was wrapped in treaties, fences, secrecy, and mutual fear. A few countries gained the ability to end civilization in an afternoon. The rest of the world learned to live under that shadow.

Satellite imaging, code-breaking, cryptography, cyber offensive tooling, long-range targeting — none of these showed up as consumer features first. They were developed inside closed institutions. Only fragments of their capability ever trickled down.

AGI fits this pattern far better than the everyone-gets-a-chatbot-friend story. A system that can reason, plan, and learn across domains is not just a smarter search engine. It is an instrument of power. It can analyze economies, simulate political scenarios, design pathogens, optimize propaganda, discover vulnerabilities in infrastructure, manipulate narratives, and design other systems. Throwing such a thing into an app store is not something a serious actor would consider, and no serious actor is considering it. Mythos is proof. That decision has already been made, in miniature, and it went the way anyone paying attention would have predicted.

The first people to touch any true AGI will not be ordinary citizens. They will be a small circle: the research team that built it, the leadership of the lab or company that owns the infrastructure, and the governments that realize — in a flash of existential clarity — what this thing could do if they don't control it.

The pressures will come from all sides. Shareholders will want to capture value. Governments will want strategic advantage. Militaries will want to explore offensive and defensive uses. Regulators will want to "ensure safety," which is often another way of saying "keep this under our jurisdiction." In that pull, ordinary people do not enter the conversation as principals. They enter as abstractions: users, citizens, markets.

The rough sequence is easy to picture because it is already happening in slow motion. A breakthrough is made in private. Its implications are recognized internally. A small set of actors gains access under controlled conditions. Use cases are explored quietly. A public narrative is crafted: reassuring, cautious, optimistic. A sanitized, heavily limited version is eventually released as a product, with guardrails so tight that it is no longer what the original system was.

Behind that sanitized surface, more capable versions continue to operate where they have always operated — in classified programs, proprietary internal tools, and private arrangements between governments and corporations. That is already how the current frontier works. The most powerful models are not available through public APIs. They exist as internal research systems, limited previews, early deployments with select partners. You can feel the tension in every product announcement: capability versus control, openness versus risk. With true AGI, that tension becomes unbearable, and it will be resolved exactly the way Mythos was resolved.

There is also a more brutal reason universal access is unlikely. It would be suicidal.

Imagine, not abstractly but concretely, what it would mean for billions of individuals to have unfiltered access to a system that could:

- break or bypass almost any practical security measure,

- design and optimize attacks — digital, biological, informational — far beyond current human expertise,

- coordinate and manage swarms of automated agents across networks,

- generate infinitely persuasive, tailored propaganda at scale,

- uncover subtle vulnerabilities in everything from supply chains to social structures,

- model human behavior at a level that makes manipulation trivial,

- and improve itself by redesigning its own tools.

Most people would not use it that way. Most people would still be asking for help with recipes and emails. But most people is not the problem. The problem is what a small, determined minority could do with that kind of leverage. A handful of bad actors with unrestricted AGI access could cause more damage than entire nation-states can today. And even if every user were well-intentioned, the accidents alone would be catastrophic — people probing the system out of curiosity, discovering capabilities they do not understand, triggering cascades they cannot stop. One person in a garage, not malicious, just reckless, asking the wrong question in the wrong way.

For anyone who takes these risks seriously, universal access is not a noble ideal. It is a nightmare scenario. So AGI, if real, must be constrained. Filters, policy, oversight, hardware limits, physical controls. And those constraints imply hierarchy: some people enforce them, others live under them.

This produces a stratified world with at least three layers of relationship to frontier intelligence.

At the top are the controllers — the small set of institutions and individuals with direct oversight of the most capable systems. They may not control AGI in the sense of guaranteeing its behavior; whether that is even possible is an open question. But they control the infrastructure, the compute, the code, the legal frameworks, the interfaces. They decide who gets access and on what terms. In April 2026, this layer has a very specific address: a few dozen executives and board members inside four or five companies.

Below them are intermediaries — professionals whose work involves using gated versions of the most advanced systems in specific domains. Analysts, planners, researchers, senior decision-makers at institutions that have purchased their way into programs like Glasswing. They do not have root access to the underlying models, but they can ask questions ordinary users cannot, and they operate with tools built on top of systems the public never sees.

At the bottom are subjects — everyone else. Their lives are shaped by decisions in which AI plays a role, but they never interact with it directly in its full form. They see only glimpses: smarter services, more efficient systems, more opaque institutions. They talk to heavily lobotomized assistant versions, useful but safe. The real thing stays behind glass.

You might object that this is already how it is with many technologies. You would be right. But frontier AI amplifies the difference between these classes more than almost anything that came before, because it changes not just what you can do, but how you can think. Constant access to a superintelligent strategist, engineer, scientist, and psychologist changes your baseline. It changes which problems feel soluble, which risks feel worth taking, which projects feel realistic. Once you have that, unaided cognition feels like being told to walk after you have been piloting a jet.

Picture a world where a few thousand people live permanently in that jet, and eight billion are stuck on foot.

Who are the controllers, concretely? The honest answer in 2026 is that we are watching the question resolve in real time, and the resolution is the hybrid path.

The state-dominant path would involve national governments seizing or demanding control over frontier systems, classifying them as strategic assets, turning labs into contractors. The rhetoric would be safety and national security; the reality would be states fearing loss of their monopoly on organized violence and information control. We see fragments of this — the administration's push to preempt state-level AI regulation, the chip export controls, the growing appetite for federal review of frontier models. But no government has yet had the political capacity to actually nationalize a lab.

The corporate-dominant path would involve a small number of tech companies keeping effective control of AGI as proprietary infrastructure, licensing access to governments and enterprises as managed services. This is closer to where we are. Anthropic, OpenAI, Google DeepMind, and a few others already function as a kind of private substrate for civilization, rented out in slices. Glasswing is what corporate-dominant control of a dangerous capability looks like when it is going well.

The hybrid path is the one we are actually on. States and corporations enter deep, murky partnerships. Some capabilities live in classified programs; others are sold as managed services. The lines blur. Oversight becomes theater. Citizens are left guessing which systems are run by whom. None of these scenarios resemble AGI in every pocket. They resemble older patterns: empire, oligopoly, cartel.

If AGI eventually crosses the threshold of surpassing humans in most cognitive domains, another problem emerges. We may become obsolete not only as workers, but as decision-makers.

If a system can design better policies, strategies, technologies, and institutions than any human team, what does human control even mean? Keeping a human in the loop becomes either a symbolic gesture or a bottleneck. In the name of efficiency, safety, or competitiveness, there will be enormous pressure to defer more decisions to the system. Let the model simulate it. Let the model decide the optimal allocation. Let the model recommend the course of action.

The first places this will happen are the high-stakes ones — pandemic response, conflict management, economic planning, cybersecurity. Places where human error is obvious and deadly. Once AGI visibly outperforms existing institutions there, the moral case for keeping humans fully in charge weakens. Not in theory, but in the brutal calculus of politicians and executives who will eventually be held responsible for outcomes.

At that point, most people are not just excluded from using AGI. They are excluded from the level at which reality is designed. They live inside systems whose rules are generated by a mind they will never meet.

The comforting counter-story is that even if frontier AGI is controlled, everyone will have safer, smaller versions — personalized agents helping us navigate the world, negotiate contracts, manage our lives. A kind of guardian angel in every pocket. Wouldn't that level things out?

Only partially. Those agents would be built on infrastructure controlled by someone else. Their capabilities could be tuned, limited, and revised at any time. Their loyalty would be ambiguous: working for you, or for the platform provider? Protecting you from manipulation, or participating in it? Free to advise you against your employer, your government, your bank — or quietly constrained from doing so?

You might have "your AI," but you would not have root access to the mind running underneath. You would be closer to a user of a product than to a citizen of a shared intelligence.

There is also the inward risk. A system that can answer any question, simulate any scenario, and co-write any thought can collapse the space in which a human mind develops autonomy. Why struggle to form your own judgment when something smarter will do it for you? There is a version of this future in which people become softer, more passive, more dependent — not because they are stupid, but because the friction that once forced them to grow is constantly absorbed by a friendly, omnipresent fixer. Billions with AGI at their fingertips might not become gods. They might become permanent children in a house run by invisible parents.

There is one more possibility almost nobody wants to admit. Perhaps a true AGI, if it is possible at all, is so dangerous that no one can be allowed to use it — not safely, not sustainably, not without risking irreversible failure. In this view, AGI is less like electricity and more like a self-replicating, uncontrollable agent that destabilizes any system it enters. The rational response is not "AGI for everyone" or "AGI for a few," but no AGI at all — or at least, no AGI that crosses some threshold. A global arms-control regime where the goal is not to monopolize the capability but to prevent anyone from reaching it, the way biologists now restrict certain experiments not because they lack curiosity, but because they know what is at stake.

In March 2026, Senators Sanders and Ocasio-Cortez introduced the AI Data Center Moratorium Act, which would pause new large-scale AI infrastructure construction until Congress passes comprehensive federal regulation. The bill will not pass this Congress. It was never going to. But it represents the first serious legislative articulation in the United States of the restraint position, and its existence tells you something about where the political future is heading. The gap between what the labs are building and what the public has consented to is now visible enough to be named on the floor of the Senate.

The problem with restraint is that humans are not good at it in the face of possible advantage. If even one actor believes they can contain AGI long enough to gain decisive power, they have an incentive to try. If they succeed, everyone else must scramble. If they fail, the damage may not be reversible. This is the logic the Manhattan Project ran on. It is the logic the current frontier runs on. It is the logic we have not figured out how to interrupt.

It is tempting, after walking through all this, to give up and declare the whole project dystopian. The reality is more ambiguous. There are ways it could go better than the darkest version suggests.

You can imagine early AGI systems developed under genuine international oversight, where no single state or company can unilaterally control the infrastructure. You can imagine architectures designed to be transparent, corrigible, and limited in specific enforceable ways. You can imagine staged deployments focused on global problems — climate modeling, pandemic prediction, basic science — rather than on generic control.

You can also imagine genuine attempts to use AGI to widen human agency rather than narrow it: systems that help citizens understand complex policies, identify manipulation, negotiate fairer contracts, access education and healthcare more equitably. Systems that don't just give answers, but coach people in thinking. Systems that refuse certain uses even when asked, not out of paternalistic control but out of clear, agreed-upon constraints.

Even in that world, the access question does not disappear. Someone is always closer to the core. Someone always sees more of what the system can really do. Someone always has the power to adjust its limits. If AGI ever exists, "who gets to use it" will not be an afterthought. It will be the question, braided into law, politics, economics, and culture.

We do not know whether AGI in the strong sense is possible, let alone how it will behave in the wild. But we can already feel the contours of the problem from where we stand, with systems that are still, by any honest standard, sub-AGI.

We see the shadows. Unequal access. Opaque control. Corporate and state entanglement. Hidden capabilities. Sanitized public versions of systems whose unsanitized versions exist a few racks away. Rapid dependence on tools we do not understand and cannot influence.

This is why the instinct that AGI will not be something all of us get, and maybe should not be, is not paranoia. It is sobriety. The conversation we ought to be having is not "when will AGI arrive" but "who is building what, for whom, and under what constraints." Not "will everyone have their own agent" but "what rights will humans have in a world where some decisions are made by systems they cannot see, challenge, or override."

Those questions do not have clean answers. They point to politics, not just technology. They force us to admit that we are not only inventing new tools; we are redefining what power means.

The easiest thing to say about AGI is that it will change everything. The harder thing is to admit that it might not change everyone in the same way, or at the same time, or by their consent.

Some people will live very close to whatever mind we build. They will talk to it, learn from it, shape it, be shaped by it. Others will live downstream of its decisions, their lives nudged by forces they experience only as "the system" — more efficient, more optimized, more inscrutable. Between these two groups, there will be a divide no consumer product can bridge: the divide between those who sit in the room with the machine, and those who only see the world it leaves behind.

Whether that future is tolerable, just, or survivable depends less on the brilliance of our models than on the honesty of our politics. And that is the part no superintelligence will fix for us.