Weapons-Grade AI and the Licensed Frontier

Frontier models are about to become the most powerful cyber weapons ever built. Governments around the world have figured this out, and their posture toward AI has changed accordingly. A model that can read code, find flaws, write exploits, and chain them together at machine speed is a weapon in every sense of the word.

In business, we talk about copilots and agents and the end of the blank page. We should also be talking about the fact that the same capability that drafts a presentation, with a different prompt and a different intent, maps the attack surface of a regional bank, a hospital network, or a power grid. Capability is capability. The model does not care what you point it at.

The Backlash Is Real and It Is Growing

At the same time, the anti-AI movement is growing. Data centers have become the physical proxy for everything people distrust about AI. Fights over water, power, noise, and tax abatements in towns most of us have never visited are now national news. Local zoning boards have become the front line of AI policy, and they are winning more often than the industry expected. The objections are not all irrational. Data centers consume real resources, and the benefits accrue somewhere else.

So, we have a real dilemma. One side of the country wants to slow this down, and it has a growing list of legitimate grievances. The other side of the equation involves adversaries who have no intention of slowing down at all.

If You Believe We Have Enemies

Every nation state on earth with the resources to compete is racing toward AGI. The ones who take it seriously are racing past AGI toward the science fiction category of ASI, artificial superintelligence. You do not have to believe ASI is imminent to understand what the race is about. You only have to believe that the next several generations of frontier models will be dramatically more capable than the current ones, and that the current ones can already do serious damage.

If you believe China is our adversary, or if you simply believe we have adversaries, the conclusion is unavoidable. Imagine a hostile state acquiring a cyber weapon powerful enough to infiltrate banks, hospitals, and power grids with intent to do harm. At a bare minimum, we would need weapons of equal quality just for defense. Cyber defense at that level is a model at least as capable as the one attacking you, running continuously, hunting for the same flaws before the other side finds them.

This puts us in a classic prisoner’s dilemma (with the game theory stripped of any subtlety). Cooperation means everyone slows down together. Defection means racing as fast as you can. No one can verify that the other players are cooperating, and the penalty for being the only cooperator is catastrophic. Under those conditions, defection is mandatory. Winning the race becomes a matter of national survival, and slowing down innovation is really not an option.

We Already Know How to Regulate Weapons

We are extremely good at regulating weapons we have decided to regulate. There are no privately owned atomic bombs. There are very few privately owned .50 caliber machine guns and even fewer privately owned functioning bazookas, at least in the United States. We built licensing regimes, manufacturing controls, chain of custody rules, and criminal penalties for physical weapons decades ago, and they work well enough that most people never think about them.

None of those regimes required us to slow down military innovation or stop building weapons. They required us to decide who gets to hold them, under what conditions, and with what accountability. That is a solvable problem. We solved it for nuclear material, for select agents in biology, and for military hardware. The frameworks are imperfect, but they exist, and they are the right starting point for weapons-grade AI.

I can imagine a near future where the most capable frontier models are regulated the way we regulate military ordnance. Access is licensed. Use is audited. The models themselves live in controlled environments with real physical and network security. The vendors who build them operate under something closer to a defense contractor’s obligations than a consumer software company’s terms of service.

The Licensed Frontier

Call it the licensed frontier. Under this model, the highest-capability systems get licensed to specific verticals with specific oversight, and the people who use them are vetted, accredited professionals with domain expertise.

Healthcare is the obvious first case. A frontier model working on cancer would be licensed to the research labs at major pharmaceutical companies and to accredited academic institutions. That is a practical requirement rather than a bureaucratic one. The model needs subject matter experts in the loop for the research to bear fruit. Oncologists, molecular biologists, and trial designers are the people who can tell a promising signal from statistical artifacts.

The same structure works in fluid dynamics, where weather prediction and aerodynamics both depend on models that are computationally brutal and physically consequential. It works in materials science, in nuclear engineering, in structural analysis, and in every other field where the ceiling on what a model can do is set by the quality of the humans in the loop. Frontier models and accredited professionals working in harmony, inside a licensed environment, is a future that respects both the capability and the risk.

Every Task Does Not Need a Frontier Model

While the frontier gets licensed, civilians will get the appropriate models for the job, and that turns out to be almost everything we actually need. After four years of deploying AI at enterprises of every size, I can tell you with confidence that most business tasks do not require a frontier model. Drafting, summarizing, classifying, extracting, translating, coding at a working level, analyzing a spreadsheet, answering a customer, and scheduling a meeting are all comfortably within reach of models that sit one or two generations behind the leading edge.

Nothing about this approach slows the race. The frontier labs keep building. The government keeps funding, and the national security apparatus gets the cyber defense it needs. The civilian economy keeps getting more productive on models that are more than good enough. And the people fighting data centers in their hometowns get something they do not have today, which is a coherent answer to the question of who is allowed to hold the most dangerous technology ever built.

I know this is not a unique viewpoint. I do not think it is an obvious future either. I think it is a probable one, and the sooner we start designing the licensing regime, the less likely we are to have it designed for us after the first real disaster.

David Sacks Responds to Dario Amodei

In a tweet, David Sacks, Chair of the President’s Council of Advisors on Science and Technology (and the former White House AI/crypto czar) offers his response to Dario Amodei’s essay about slowing the pace of the frontier. These are David’s words, and I think he’s spot on:

Dario has written that we need to “pace the frontier,” and Sam has agreed. People may be surprised by my response: go ahead.

You guys are the frontier. By any reasonable metric — market share, revenue growth, model capability — the two of you have a duopoly on frontier intelligence. You’ve also claimed the lead is widening because of recursive self-improvement.

I don’t see what you see in the lab. If the unreleased models are scary enough that you think you should slow down, I support your decision to be responsible.

But stop pretending you need anyone else’s permission. Stop pretending antitrust law has to be suspended so you can form a cartel. Stop pretending you need a regulatory approval process that supersedes product liability. Stop pretending METR is independent when it is intertwined with Anthropic’s investors and staff. Stop pretending you need those same evaluators to police competitors who aren’t even at the frontier.

Most of all, stop pretending the motivation to slow down is purely altruistic. You face massive product-liability exposure if your products enable a truly damaging cyberattack. The market already punishes models that behave in unpredictable or unauthorized ways. After the Hugging Face episode, it is simply good business for OpenAI and Anthropic to trade some raw power for reliability and predictability. Call it alignment if you want. It is also just giving customers what they want.

Pacing the frontier would also create breathing room for a more intelligent conversation about regulation than Bernie Sanders’ “shut it all down.” China is very unlikely to join a global agreement, as you know, and that has to be taken into account as well.

So go ahead and pace the frontier. You are the ones setting it. The easiest way not to build superintelligence is for you to agree not to build it. Demanding your preferred regulatory framework as the price of that will look like blackmail of the public and the political system. So just do it.

If you do, you’ll buy goodwill for the next conversation. If you don’t, we’ll know this was just another bid for regulatory capture — or an election-season psyop.

Author’s note: This is not a sponsored post. I am the author of this article and it expresses my own opinions. I am not, nor is my company, receiving compensation for it. This work was created with the assistance of various generative AI models.