The US just reversed itself on one of the odder tech-policy experiments in memory: an export freeze on an AI model. The government has lifted the controls it had placed on Anthropic's most advanced systems, restoring foreign access, the BBC reported — ending a weeks-long episode that revealed just how awkwardly old export rules fit modern AI.

What happened

In early June, Anthropic — the AI company behind the Claude models — released two frontier systems (reported as Mythos 5, aimed at advanced cybersecurity work, and Fable 5 for broader use). Soon after, the US Commerce Department imposed emergency export controls, citing concern that the technology's power to find security vulnerabilities could aid cyberattacks in the wrong hands. The restriction effectively barred non-US users — including, awkwardly, Anthropic's own foreign staff and overseas customers — from the new models, CNN reported.

The government then eased the block (clearing a set of approved companies), and on June 30 lifted it entirely. Exactly which countries and customers regain full access, and under what conditions, isn't fully spelled out — treat the scope as still being clarified.

The real lesson: rules for chips don't fit software

The episode exposed a genuine problem. Export controls were designed for physical things — machines, weapons, and lately the advanced chips that power AI — that you can stop at a border. But a frontier AI model lives on servers and is used remotely over the internet. There's no crate to inspect. You can't easily wall off "foreigners" from software delivered through the cloud without, in effect, shutting the product down — or firing your own overseas engineers.

In short, Washington tried to treat an AI service like an exported weapon, and the mismatch showed immediately.

Security vs. market access

Behind it sits a real tension the US keeps wrestling with. On one side, national security: the government has spent years restricting advanced AI and chips to keep cutting-edge capability away from rivals like China, on the logic that the same AI that boosts productivity can also supercharge cyberattacks. On the other, commercial reality: US AI firms argue that heavy-handed limits cost them global customers and push the world toward Chinese alternatives — ceding the race to sell AI "the American way."

The quick reversal suggests officials weighed those commercial stakes heavily. It also fits a broader remaking of US AI-export policy that has swung between loosening old rules and intervening directly in specific companies' access — a less predictable, case-by-case approach.

Why it matters

For Anthropic and its US peers, the whipsaw is a reminder that regulatory risk — not just product quality — now shapes growth: access to foreign markets can be switched off overnight. For the AI industry Boursel tracks — from Nvidia's chips to Micron's memory to the data-center build-out — it underscores that geopolitics is now woven into the business. And for policymakers, it's an early, messy sign that governing AI will require new tools, because the old ones — built for a world of shipped goods — don't cleanly apply to intelligence delivered down a wire. Boursel takes no side on the policy; the takeaway is that the US just learned, in public, that you can't embargo a cloud the way you embargo a chip.