The Breach of Trust

At 5:21 p.m. Eastern time on June 12, 2026, a letter arrived. A few hours later, two of the world's most capable AI models — Anthropic's Fable 5 and Mythos 5 — were gone. Not for some malicious minority, not for a named adversary, but for every foreign national on earth. A Swedish engineer, a Norwegian researcher, a German factory: all of them suddenly facing the same emptiness. The tool they had built their working day around no longer existed.

The official reason was national security and an export-control order. The alleged threat was a method for bypassing the model's safeguards. The provider itself, Anthropic, disagreed: it complied with the order but called it a misunderstanding, arguing that the vulnerability invoked was trivial and just as available in other models. Who is right on the technical merits, we do not yet know. But that is not where the real event lies.

The real event is that the switch turned out to exist — and that it is not in your hand.

We have grown used to thinking of frontier models as infrastructure. Like electricity in the wall, water in the tap. Always there, always on. But electricity and water are regulated as essential utilities under laws you can shape with a vote. A frontier model from an American company is something else. It is more like the services of a foreign embassy: available as long as the relationship between two powers permits, and closable overnight without anyone asking you. What was "exported" that night was never the model. It was the knowledge that the off-switch exists, and that it belongs to someone else.

This is what I want to call a breach of trust — not necessarily a betrayal by the provider, who in fact protested, but a structural breach. Trust in a dependency is never stronger than the weakest sovereign in the chain. The Swede, the Norwegian, the German who had built their workflows on the model never had a voice in the room where the decision was made. They were not really trusting the technology. They were trusting that the relationship between Washington and its own security apparatus would remain stable and benign. That night, the trust proved misplaced — not in its strength, but in its object.

And trust of that kind does not return unchanged once broken. You can regain access — Anthropic is already working on it — but you cannot regain the innocence. The question has changed. It is no longer "which model is best?" but "which model cannot be taken from me?" That is an entirely different calculus, and it does not favour whoever owns the switch.

Here it is tempting to draw a triumphant European conclusion: therefore, our own models; therefore, sovereignty. But honesty demands an objection. The alternative is not clean. Open weights you run yourself are something no one can switch off for you — but they often trail a step behind, and the word "sovereignty" can become a comforting story that hides a capability gap. Europe will not build something to beat the American frontier models overnight. What has changed is not capability but calculation. The price of dependence has become visible, and visible prices are eventually paid.

The night a model went dark will be remembered less for what was lost — a temporary suspension of a single family of models — than for what it revealed: that the light was never ours to keep on. And every builder who woke to that emptiness now faces a question that can no longer be deferred: what are you willing to build on something that someone else can switch off?

Sources and checks

The sources below cover the core factual claims about the export-control directive, timing, Anthropic's objection and the effect on Fable 5 and Mythos 5. The conclusion about trust and sovereignty is the author's.

Rolf Skogling writes AI-skiftet from an industry-near and practical perspective, grounded in work with AI in real organisations.