There is a particular kind of decision that tells you more by its manner than its content. When the US Commerce Department gave Anthropic an hour and a half on a Friday to take down Fable and Mythos — the flagship models of a company valued only days earlier at $900bn — the haste was the message. You do not give ninety minutes’ notice to a firm you are gently nudging. You give ninety minutes to a firm you wish to discipline.
The official account is tidy enough. Researchers at Amazon found a way to coax Fable past its guardrails into naming software vulnerabilities. Andy Jassy carried the finding to the White House; ministers conferred; the president signed off; export controls followed, barring foreign nationals and, by extension, forcing Anthropic to pull the plug for everyone. National security, the administration says — nothing personal.
It is when one examines the instrument that the story begins to fray. Export controls exist to keep sensitive technology out of the hands of adversaries. Yet the practical effect here is the opposite of containment. The capability Amazon demonstrated is not unique to Anthropic; by the account of researchers across the field, OpenAI’s and Google’s models can do much the same — and those models remain untouched. So the rule does not deny the capability to anyone determined to have it. What it denies is access to allies: Britain, the European Commission, the very partners who had been using Anthropic’s Mythos to find and patch holes in their own networks.
“The administration got it wrong,” said Katie Moussouris, the cyber-security veteran who read Amazon’s report. The guardrails, she found, were working as intended; there was, in her phrase, nothing to patch.
Pile the contradictions one atop another and the structure looks less like policy than improvisation. The same department that issued the ban had tested and cleared Fable for release days earlier. The vulnerabilities the model surrendered were minor and already known — the sort of thing a network defender would happily prompt for. And while Commerce moved to control a model that allegedly imperils the nation, the same administration has been loosening restrictions on the advanced chips that build such models, blessing shipments to China. As one analyst dryly noted, it is a curious posture to police the recipe while exporting the oven.
Helen Toner, formerly of OpenAI’s board, put the technical objection plainly: jailbreaks cannot be fully fixed — it is an inexact science — and any government should assume rival models are capable of the same tricks. Her verdict on the chosen lever was sharper still.
Export control, she argued, is poorly suited to the task; it bars foreign governments from a tool precisely when allies are meant to be wielding it for cyber-defence. The European Commission, newly granted access and now contemplating its withdrawal, was diplomatic but pointed: contingency measures, it said, should not discriminate against partners.
So if the ban does not contain the danger, does not fix the flaw, and wounds the friends it claims to protect, one is entitled to ask what it is for.
Here the longer feud comes into view, and with it a more coherent reading. Anthropic and the Trump Administration have been at odds for months — over the guardrails governing the Pentagon’s use of its models, over lethal autonomous weapons, over domestic surveillance.
The Pentagon has branded the company a supply-chain risk; the two are in litigation over the label. Defence Secretary Pete Hegseth, never one for understatement, boasted that his department had thrown Anthropic “out of our building — forever,” adding that each passing day vindicated the move. The venture capitalist David Sacks has accused the firm of “doomerism” — of frightening the public about AI’s perils while shipping ever more powerful systems. There is, in other words, no shortage of grievance in search of an outlet.
Then there is the matter of timing. Anthropic had just leapfrogged OpenAI to become the world’s most valuable AI start-up, with a $47bn revenue run-rate and an IPO in prospect. A ninety-minute order that darkens your flagship product is not merely an inconvenience to such a company; it is a demonstration — to investors, to rivals, to the firm itself — of who holds the switch. Read this way, the episode is less an export control than a show of sovereignty over frontier AI, dressed in the borrowed clothes of a cyber-emergency.
The casting adds its own intrigue. Amazon, which sounded the alarm, is also a $13bn investor in Anthropic; Jassy is said to have raised broader worries about frontier capability rather than this single flaw. A patron quietly flagging its protégé to the state is the sort of detail that rarely survives into the official telling.
Step back far enough and the whole exercise acquires a quixotic glow. For while Washington frets over the singular menace of one American model, the market is moving in precisely the opposite direction — towards abundance. A price war has broken out; firms are routing all but their hardest tasks to cheap open-source systems, many of them Chinese, at a fraction of the cost. Intelligence is becoming a commodity, not a crown jewel. To sandbag a single model in this flood is to mistake a stream for the river. The capability the administration wishes to bottle is already pooling everywhere, and getting cheaper by the month.
Which brings us to the moral of this particular Fable. A company that made its name warning — loudly, and lucratively — about the dangers of artificial intelligence has been punished with a danger it insists is illusory: its own cautionary tale turned into a cudgel. Perhaps the administration genuinely believes it has averted catastrophe. But on the evidence so far, the ban contains nothing, fixes nothing, and protects no-one save the prerogative of those who imposed it. That may, in the end, have been precisely the point.
• The writer is Deputy Managing Editor, Gulf Times.