Aesop's Fable 5

I couldn’t help but be reminded of Aesop upon reading that Anthropic had, on Friday, begun switching off its two most capable models.

The fable is one which, if you don’t know, you will undoubtedly know its lesson. It is called The Eagle and the Arrow. An eagle sits on a high rock, scanning the ground for a hare, when a hunter’s arrow strikes it. As the bird lies dying it turns its head to the shaft that killed it and sees, fixed into the wood, one of its own feathers. The arrow could not have flown straight without it. The lesson the Greeks drew was a hard one: we so often hand our enemies the means of our own destruction.

This week, Anthropic played the eagle. When it switched off Fable 5, the public-facing version of Mythos, it was complying with a government it had spent years arming.

On Friday the Trump administration issued an export control directive, citing national security, barring any foreign national from touching the models anywhere on earth, employees included. The net effect was that Anthropic had to disable both Mythos 5 and Fable 5 for all customers, including those within the US. This is arguably the most aggressive step Washington has yet taken against a frontier model.

Whilst the directive named no specific reason, Anthropic's own reading is that it stems from a single, unconfirmed method of jailbreaking Fable, which the company says is no grounds for recalling a model deployed to hundreds of millions of people. Reporting since points the finger at Amazon, Anthropic's largest investor, whose chief executive told Treasury Secretary Scott Bessent that its own researchers had drawn cyberattack material out of Fable, the precise danger Anthropic had spent months advertising.

Governments are slow. They do not, as a rule, work out from first principles that a software product is a strategic weapon to be pulled from the market under national security law. They are told, and then, eventually, they act. This was the tail wagging the dog. The judgement that a model could be dangerous enough to seize was supplied, and the industry has spent the past years doing the supplying.

Consider how the arrow was made. Anthropic released Mythos in April to a closed circle of government and corporate partners under a programme it called Glasswing, withheld from the public on the grounds that it was too dangerous to hand over freely. It briefed those partners on the model’s gift for finding live cybersecurity flaws. When the consumer version followed last week, it came sheathed in guardrails against cyber and biological misuse. All of this was sincere. Anthropic believes it is the responsible lab and behaves accordingly. But every warning was also an advertisement. If you spend years telling the government of the most powerful nation on earth that you are building something uniquely dangerous, it should come as little surprise when it agrees with you.

Dario Amodei has made the nuclear comparison his house style. He has called the export of advanced chips to China the equivalent of “selling nuclear weapons to North Korea.” He has written that frontier models will soon resemble weaponisable nuclear material rather than aircraft or cars. In a long essay on AI and the state this week, he argued that governments should hold the power to block or reverse the deployment of any model that poses unacceptable risk.

The directive that blocked Fable is the authority he asked for, exercised against him. Yet none of this has come from nowhere. The same administration spent the spring stripping Anthropic’s models out of federal agencies, after the company refused to drop its limits on mass surveillance and autonomous weapons. The appetite for a fight was already there. The danger talk did not create it, and was most likely an attempt to stymie it. What it supplied instead was the justification, and a government hunting for leverage rarely turns one down. Compare your product to a nuclear weapon for long enough and the Eye of Sauron turns your way. When it does, it does not negotiate. It requisitions or removes it entirely.

Nobody issues an export control directive against a project management app. The state keeps that machinery for enriched uranium, for encryption, for satellite imaging. On Friday it told us, again, which shelf frontier AI sits on.

Sam Altman has spent the year describing intelligence as a utility, something the public will “buy from us on a meter” like electricity or water. He should study his own comparison. Electricity became a utility, and precisely because it did, it stopped being a thing a private company simply owned. It became the most heavily regulated industry on earth, and across most of the world, at one time or another, the state took it outright.

The deeper exposure, of course, goes further than Anthropic. A single government reached into a product held by hundreds of millions of people, in every jurisdiction on earth, and switched it off. No other government was consulted, and none was in any position to protest. A model you do not control is not owned but lent; and a thing that is lent, as every borrower eventually discovers, is recalled at the lender’s convenience and never at your own.

In any case, when it comes to this technology: pick your metaphor. AI as a nuclear weapon. AI as a utility. The first is owned by states. The second is regulated and, in much of the world, nationalised by them. Name your work a weapon often enough and the state will take you at your word. Name it the foundation of the future and you have told the state precisely what it cannot afford to let you keep.

Every warning is a feather and every promise a shaft, and there is only one archer collecting them. The labs still in the air should think hard before plucking the next one.