On June 9, 2026, Anthropic launched Fable 5, the most powerful model it had ever made public.
On June 12, a directive from the US Department of Commerce ordered it to cut access for all foreign nationals — including its own non-American employees. Unable to verify each user's nationality in real time, Anthropic cut both models for everyone.
On June 30, the controls were lifted. On July 1, global access was restored.
19 days. No trial. No negotiation. No vote. An on/off switch operated from Washington.
What actually happened
Let's get the facts straight, because they deserve precision:
- June 2, 2026: President Trump signs an executive order asking AI companies to "voluntarily" submit their frontier models for a government security review before public release.
- June 9: Anthropic launches Fable 5 (the safeguarded consumer version) and Mythos 5 (a less restricted version of the same underlying model).
- June 12: the Commerce Department's Bureau of Industry and Security imposes export controls on both models, citing national security. Anthropic must cut access immediately.
- June 30: the controls are lifted. Fable 5 becomes globally available again on July 1. Mythos 5 remains restricted to government-approved American organizations.
And the story doesn't end with Anthropic. On June 26, OpenAI launched GPT-5.6 — available only to a small group of partners approved customer-by-customer by the White House. Sam Altman himself told the government this was "not our preferred long-term model." OpenAI complied anyway.
In three weeks, government access control went from hypothesis to standard procedure. The executive order calls the review "voluntary"; in practice, no American lab now ships a frontier model without Washington's green light.
The real problem isn't the cutoff. It's that it's possible.
Access has been restored, and that's almost worse: it creates the impression the episode is over. It isn't.
What those 19 days demonstrated is a mechanism:
- Your AI provider is subject to US law.
- The US government can order it to cut access, immediately, for reasons it alone defines.
- The provider complies — Anthropic did exactly what it was told, and you can't blame them.
- You have no recourse. Not contractual, not legal, not commercial.
If your company has built Claude or GPT into its processes — customer support, document production, analysis — the question is no longer "could this happen?". It happened. The question is: how long does your business survive if it happens again, for longer?
This is not an anti-American position. It's the same reasoning we apply to the Cloud Act for data: what matters is not the provider's intent, it's the legal structure it operates in.
Europe has already started drawing conclusions
In April 2026, the European Commission awarded a €180 million, six-year contract to host EU institutions' data — exclusively with European providers: Post Telecom with OVHcloud and CleverCloud, StackIT, Scaleway, and Proximus with Mistral. American hyperscalers were not selected directly (one nuance: the Proximus lot relies on S3NS, a Thales–Google Cloud joint venture operated exclusively by EU companies).
When the Commission, the Council, and the ECB decide their own data will no longer run on infrastructure subject to US law, it's no longer a paranoid fantasy. It's infrastructure, signed contracts, committed euros.
What this forces you to rethink in your AI architecture
Concretely, three principles protect a company from the kill switch:
1. Models nobody can take away from you. Open-weights models (Gemma, Qwen, Mistral…) are downloaded and run on your side. Once the weights are on your server, no administration can remotely disable them. Our 2026 open-source model benchmark shows quality and speed are no longer behind proprietary APIs for most business use cases.
2. A multi-model architecture, without lock-in. If your processes are built on a single model, you inherit all of its risks. A platform that routes across several models — including at least one running inside your perimeter — turns a cutoff into a temporary degradation instead of a full stop.
3. Infrastructure under European law. European operator, European hosting, or on-premise deployment: the same principle as for the Cloud Act. Protection holds by construction, not by contract.
That is exactly Eridia's architecture: a multi-model platform, open-source models running on your infrastructure or hosted in France, operated by a French company. On June 12, our clients noticed nothing.
The question to bring to your next board meeting
The debate isn't technical, it's political and strategic: does your company accept that its AI infrastructure has an off button controlled by a foreign government?
If the answer is no, the exercise comes down to one question for every provider in your AI stack: "if Washington cuts access tomorrow, what keeps running on our side?"
And if you want a clear picture, we offer a free assessment of your exposure: which tools, which models, which dependencies, and what should be brought home.
