Regulating the Frontier Without Closing the Door

On Tuesday last week, Dario Amodei, the chief executive of Anthropic, published an essay called Policy on the AI Exponential. Its central proposal was that the most powerful AI models should face mandatory third-party safety testing, and that the government should hold the power to block or reverse the deployment of a model that presents unacceptable risk. He drew the analogy to aviation: rigorous oversight at the top, so that safe and routine access can flow to everyone below.

Three days later, on Friday, at 5:21pm Eastern Daylight Time, the United States government ordered Anthropic to suspend access to its two most capable models, Fable 5 and Mythos 5, for every foreign national across the world. And so, within hours, the company disabled both models for all of its customers worldwide. The power Amodei had argued governments should hold was exercised against his own company, only three days after he argued for it.

I had already drafted an article discussing that essay, when this happened. My argument was hopeful: that as AI governance hardens at the top, the thing worth protecting is broad access at the bottom, because access is what has quietly begun to level the professional playing field in New Zealand. I still stand by that argument. But the week overtook it. The question is no longer whether the door at the bottom might one day be narrowed. It is that the door was closed, without warning, on a Friday evening, for everyone outside one country. This adds a degree of urgent concern. Leaving access in the hands of hope is clearly naive.

What actually happened

The facts, drawn from Anthropic's own statement and from independent reporting, are as follows. The US government issued an export control directive citing national security authorities. It required Anthropic to suspend all access to Fable 5 and Mythos 5 by any foreign national — whether inside or outside the United States, including Anthropic's own foreign-national employees. Because selective compliance would have meant identifying and blocking a vast swathe of users, the company chose to disable both models entirely, for everyone. Access to Anthropic's other models was not affected. The stated trigger was a reported method of bypassing, or "jailbreaking", one of Fable 5's safeguards. Anthropic says it reviewed a demonstration of the technique and found that it surfaced only a small number of previously known, minor vulnerabilities. The letter, reportedly issued through the Commerce Department, did not set out the specific national security concern in detail.

None of this happened in a vacuum. Earlier this year the United States Department of Defense had designated Anthropic a supply chain risk, and the company is in continuing litigation against the administration. Whatever else this is, it is not a neutral technical safety review. It is the exercise of state power over a commercial product, in a contested relationship, with national-security framing and little public detail.

The objection that matters

Anthropic complied. But its public response drew a careful and important line. The company said it disagrees that the discovery of a narrow potential jailbreak should be cause for recalling a commercial model deployed to hundreds of millions of people, and warned that if that standard were applied across the industry it would effectively halt all new model deployments by every frontier provider.

Then it said something worth dwelling on. Anthropic restated it believes the government should have the power to block unsafe deployments — but only as part of a statutory process that is, in its words, transparent, fair, clear, and grounded in technical facts. This current action, it said, did not adhere to those principles.

The fault line was nationality

My original draft concerned itself with access being narrowed by capability tier — the most powerful tools flowing to large institutions while individuals waited. That can still happen. But that is not what happened this week. This week, access was cut along a different and starker line: nationality. The directive did not distinguish between a careful enterprise and a careless one, or between a safety researcher and a bad actor. It distinguished between Americans and everyone else.

For a New Zealand practitioner, that line is not abstract. I am a foreign national. So is every lawyer, every HR manager, and every employer in this country. The most capable publicly released model from the world's most safety-focused laboratory became, for an evening and for all of us outside the United States, simply unavailable — by order of a government none of us elected, for reasons none of us were told, with no notice and no process we could engage. The "nuclear outcome" I had described in the abstract — access gated through national-security machinery — did not arrive as a distant possibility. It arrived as a Friday-evening email.

It may well be temporary. Anthropic has said it believes the action rests on a misunderstanding and is working to restore access. I hope it succeeds, and quickly. But for the purpose of governance, whether it is reversed is beside the point. What matters is that it is now demonstrated. Frontier-model access for non-Americans has been shown to be contingent, revocable overnight, on national-security grounds, without warning. A thing that has happened once cannot be treated as something that merely might happen. It is a known property of the system now.

What this did to the firms that had already built on it

The individual practitioner losing access for an evening is the small version of the story. The larger version belongs to the organisations that had already wired these models into their workflow systems. In the few days between launch and suspension, enterprises had begun integrating Fable 5 and Mythos 5 into software development, research, and business automation. When the models went dark, those integrations went dark with them.

The lesson the infrastructure community drew is the one that travels furthest: frontier-model availability can become a compliance event, a vendor-risk event, and a production-continuity event simultaneously. A single regulatory letter, addressed to a vendor in another country, can stop a tool that an organisation has come to depend on, not because the organisation did anything wrong, but because of where it, or its people, happen to be from.

Now bring that home to a New Zealand law firm. A growing number of firms run their practice on software built on top of frontier models — research tools, drafting assistants, review platforms, often costing thousands of dollars a month and frequently powered by Anthropic's models specifically. Many of those firms could not tell you, today, which underlying model their vendor uses, or what happens to their workflow if that model is suddenly withdrawn. This week supplied the answer for those who were on the newest model: the work stops. Matters in progress stall. The drafting assistant that a fee earner had built a Friday afternoon around is simply gone, and there is no one local to call, because the decision was made by a foreign government about a foreign vendor.

That is not a hypothetical risk to be managed at leisure. It is a continuity exposure that crystallised just last week, and it sits squarely inside obligations New Zealand lawyers already carry. A firm that cannot deliver because a critical tool vanished without notice still owes its clients competence and diligence. The professional answer is not to abandon these tools — they are too useful, and the levelling they offer is real — but to hold them at arm's length the way a prudent firm holds any critical third-party dependency: know what you are running on, know what you would do if it disappeared, and never let a single externally controlled model become a single point of failure for client work.

A principle that travels further than its author intended

Return to the distinction Anthropic drew — not against the power, but against the process. The company's objection was that the directive lacked transparency, was not grounded in clear technical facts, and gave those affected no process to engage with or respond to. Strip away the subject matter and what remains is a principle that any New Zealand employment practitioner will recognise immediately: that the difference between a legitimate exercise of power and an arbitrary one is procedural in character. Not whether the power exists, but how it is used. Not whether the outcome was justified, but whether the affected party had any real opportunity to be heard before it was imposed.

That principle sits at the heart of good faith under s 4 of the Employment Relations Act 2000, and of the justification test under s 103A. Both turn on process — on whether a fair and reasonable employer acted openly, on the facts, and gave the affected person a genuine opportunity to respond before the decision was made. The argument Anthropic advanced this week about what responsible state regulation of AI should look like is, in its procedural structure, the same argument New Zealand employment law has been making about workplace decisions for twenty-five years. It is a familiar argument in an unfamiliar setting, and it is the bridge between this story and the daily work of New Zealand practice.

Where this leaves New Zealand employers and firms

Three things to carry out of this week. The first is that regulation of the model is still not governance of its use. Even now, with the frontier-control question vivid and unavoidable, nothing about it relieves a New Zealand employer of governing its own AI use. A model that passes every safety test, or fails one and is pulled, is equally incapable of making your redundancy consultation defensible under s 103A. That remains entirely your responsibility, and it is the part you actually control.

The second is that access is now a governance variable, not a given. When you evaluate an AI tool, or the contract behind it, the durability of access has become a real question to ask: what is this built on, who can switch it off, and what is our position if they do? A workflow resting on access that can be revoked by a foreign directive is a dependency to be managed, not an asset to be assumed.

The third is that the hope I began with survives, but it has grown up. The aviation outcome — rigorous oversight at the top, broad and routine access below — is still the outcome worth wanting, and still achievable. But this week showed that it is not the default. Access can be cut along lines that have nothing to do with safety and everything to do with sovereignty, and small countries full of foreign nationals are on the wrong side of those lines by definition. If we want the open door, we will have to be clear-eyed about who holds it, and build our practices so that no single hand can close it on us without our having a plan.

The door worth holding open

A disclosure, because a governance practice should model what it preaches: Lex Praxis is AI-augmented, and the AI it is augmented with is Claude, built by Anthropic — the company at the centre of this story. Every output here is reviewed and owned by a practitioner under a documented verification framework. I have tried to report the week fairly and from independent sources; readers can weigh the connection as they see fit. I will say plainly that I read the government's action as heavy-handed and the company's objection as sound, and that this was, by every account including the government's own framing, something done to Anthropic, not by it.

Amodei opened his essay with Treebeard, Tolkien's ancient tree-shepherd — wise, but so slow that greeting a neighbour takes a full day — as a figure for institutions confronting a technology that moves at exponential speed. His worry was that the trees move too slowly. This week revealed the other half of the problem. When the trees finally do move, they can move very fast indeed, in directions no one anticipated, and a great many people standing in the shade can find the light cut off between one Friday afternoon and the next. The frontier will be governed. The only real questions are by whom, by what process, and whether the rest of us — our voices, outside the room, outside the country — are left any standing to be heard.

Technology evolves. Governance must keep pace. And this week, governance moved, just not in the way, or from the direction, that anyone writing about it on Tuesday expected.