In June, the US government ordered Anthropic to suspend access to Fable 5 and Mythos 5 for foreign nationals, including those inside the United States. Unable to verify everyone’s nationality immediately, Anthropic switched the models off for all users. Washington had decided who could use some of the world’s most capable artificial intelligence. The rest of the world could wait.
The restrictions were lifted on 30 June, and Fable returned globally the following day, with Mythos access restored more selectively. But the reversal did not erase the demonstration of power. If advanced AI becomes central to economic life, a decision taken in Washington could interrupt productive capacity across countries whose citizens had absolutely no say in it.
The summer also supplied reasons to take AI’s dangers seriously. On 30 July, Anthropic disclosed three incidents in which models conducting cybersecurity tests had gained unauthorised access to real systems. Internet access had mistakenly been left open. Its August follow-up acknowledged operational and behavioural failures, described pauses and stronger safeguards, and supported coordinated restraint.
These were real failures requiring real remedies. But there is a dangerous slide from “this technology needs better controls” to “our country should decide who gets it”. A passport check does not repair an unsafe testing environment. Nationality is not a technical safeguard, however reassuring it may look on an authorisation form.
By September, Anthropic was describing an access programme developed with the US government for Mythos’s advanced biological capabilities. Restricting dangerous biological assistance can be justified. The much larger question is what happens if selective access becomes the organising principle for advanced intelligence itself. Who gets the best systems, who gets something weaker, and who is expected to accept that this is for their own good?
Imagine three companies developing the same software product. With the best AI, the American company finishes in twenty-four hours. The European company needs a week. The Iranian company needs a year. These are hypothetical figures, but the consequences are straightforward. The American firm can improve and sell its product while its competitors are still building theirs. The Iranian firm may finish just in time to discover that its market no longer exists.
Now suppose the European and Iranian firms must pay through the nose for their inferior tools. The suppliers keep the strongest capabilities, charge foreigners for something weaker and use the revenue to develop the next generation. You would be paying your competitors for permission to remain behind them. Your firms earn less, your country has less to invest, and your payments help finance the widening gap.
Extend that across engineering, medicine, manufacturing and scientific research. We are talking about livelihoods, public revenue, functioning health services and the ability of whole societies to improve their circumstances. A hierarchy of AI access could become a hierarchy of national prospects. Billions of people could find their opportunities constrained by decisions made to preserve somebody else’s strategic advantage.
The nuclear comparison should make us more demanding, not more accepting. Non-proliferation can reduce genuine danger while preserving the privileges of countries that already possess the bomb. But forgoing nuclear weapons does not mean accepting that your civilian industries must work at a fraction of their competitors’ speed. AI restrictions could reach into almost everything a country produces.
The Nuclear Non-Proliferation Treaty at least offers reciprocal commitments involving peaceful nuclear cooperation and negotiations towards disarmament. Where is the equivalent bargain if AI’s leading powers reserve the best intelligence while selling everyone else a subscription to second-class status? Access to yesterday’s model is a miserable concession if today’s model is what makes your competitors unbeatable.
Nor is this merely governments misunderstanding researchers. Anthropic had already argued in May that a commanding American lead would improve the prospects for safety cooperation with China. First we secure superiority; then we can afford restraint. Unfortunately, the follower cannot afford to remain behind and the leader cannot afford to let it catch up. Every warning about the dangers becomes another reason to win the race.
Permanent exclusion may also prove impossible. Open-weight models are already distributed worldwide and cannot reliably be recalled. They may help researchers build better successors, although we do not yet know whether today’s models can support that process all the way to the frontier. Computing resources remain a constraint. Still, restrictions may buy time rather than close the route, giving excluded countries every incentive to develop independently and conceal what they are doing.
That is an appalling potential outcome: an increasingly unequal world, an accelerating technological race and less visibility into the work most urgently requiring scrutiny. Calling the policy “safety” would not make those consequences disappear. It would merely give the countries benefiting from it a more comfortable account of their behaviour.
Serious safety arrangements must constrain the powerful as well as the excluded, and preserve a credible route to civilian prosperity. Otherwise, we risk borrowing the nuclear order’s division into haves and have-nots and extending it across the world economy. Non-proliferation, AI edition: the powerful keep the means of progress, everyone else pays rent, and we congratulate ourselves on protecting humanity.