At 5:21 PM Eastern Time on Friday, June 12, 2026, a letter arrived at Anthropic’s offices from US Commerce Secretary Howard Lutnick. It cited national security but did not explain the threat in detail. Instead, it instructed Anthropic, under decades-old export control law never before applied to a commercial AI model, to immediately cut off every foreign national on Earth — including its own employees — from its two newest and most capable models, Mythos 5 and Fable 5. Because there is no way to verify a user’s citizenship through an API call, Anthropic had one option: disable both models for every customer, everywhere, instantly. This AI model shutdown is the full, carefully verified account of what happened, why three different explanations are circulating, and what it means for any organisation that has built its operations on top of a frontier AI model it does not control.

This AI model shutdown is not a story about a cyberattack, a data breach, or a model malfunction. It is a story about power — specifically, the previously theoretical question of whether a government can switch off a commercially deployed AI model that hundreds of millions of people rely on, and what happens when that theory becomes Friday-evening reality. Consequently, understanding this episode in detail is now essential context for any enterprise technology leader making AI vendor decisions in 2026.

How a Single Letter Triggered an AI Model Shutdown Worldwide

The Export Control Behind the AI Model Shutdown

The legal basis for this AI model shutdown is, on its own, a significant story. According to expert analysis from the Center for Strategic and International Studies, Lutnick’s letter invoked the Export Control Reform Act of 2018 through what is known as an “is-informed” letter — a mechanism that lets the Commerce Department privately notify a company that a license is now required for an emerging or foundational technology. Crucially, there is no existing regulatory framework in the Export Administration Regulations implementing this authority for AI models. It had never been used this way before.

In practical terms, the letter required an approved export license for sending the model outside the US, moving it between foreign countries, retransferring it within a foreign country, or releasing it to any “foreign person” — whether that person is sitting in Singapore or sitting in Anthropic’s own San Francisco office. Furthermore, since API authentication has no reliable way to verify a user’s citizenship in real time, full compliance left Anthropic with only one option: a worldwide shutdown of both models, regardless of where any individual customer happened to be.

Anthropic’s Case Against the AI Model Shutdown

Anthropic’s public statement laid out a specific technical rebuttal. According to the company, the government’s concern centred on a demonstrated method of “jailbreaking” Fable 5 — but Anthropic characterised the actual technique as trivial: “asking the model to read a specific codebase and fix any software flaws.” The vulnerabilities surfaced this way, Anthropic said, were “previously known” and “minor,” and the same capability “is widely available from other models (including OpenAI’s GPT-5.5).”

Moreover, independent technical opinion broadly supported this framing. Katie Moussouris of Luta Security — reportedly the only outside security expert to review the underlying research in detail — said researchers had fed the model open-source code containing known and deliberately planted flaws. When directly asked to review the code “for security issues,” the model declined. When asked instead to simply “fix this code,” it produced patches that could, with manual extra steps, be converted into usable exploit scripts. Therefore, she called this “defensive prompting,” not a genuine bypass of safety training — and noted that this behaviour “cannot meaningfully be fixed” without weakening the model’s core usefulness for the very defensive work Project Glasswing was built around.

Independent Security Researchers Back Anthropic’s Account

An open letter organised by Alex Stamos, the former chief security officer at Facebook and Yahoo, grew from 76 to more than 150 signatories within days. It argued that Mythos-class models are “not uniquely good” at finding vulnerabilities relative to other frontier models, and that removing them from circulation actively harms the defenders who had been using them to patch software ahead of attackers.

Why the AI Model Shutdown Happened — Three Accounts

No single confirmed account fully explains why this AI model shutdown landed when it did. Three separate narratives have emerged from reporting, each with a different evidentiary strength. Straithead presents all three, clearly labelled by how well-sourced each one is.

Click Each Account to Read the Full Evidence

An AI Model Shutdown for Allies, Chip Exports for China

Furthermore, the directive’s optics became considerably more awkward when set against another concurrent policy decision. Earlier in 2026, the same administration agreed to let NVIDIA and AMD sell advanced AI chips — the H20 and MI308 — into the Chinese market, with the US government taking a 15% cut of related China revenue as a condition of approval. Former Trump AI adviser Dean Ball called the chip decision “baffling” and “cartoonish.” Consequently, critics seized on the apparent inconsistency immediately: Washington permitted advanced chip hardware to flow toward a strategic rival while simultaneously blocking treaty allies — Britain, France, Canada, Japan — from a frontier US software model.

What This AI Model Shutdown Means for Concentration Risk

Set aside, for a moment, which of the three explanations for this AI model shutdown is correct. The structural fact that matters for every enterprise technology leader is this: a frontier AI model serving production workloads at hundreds of organisations was disabled, globally, within hours, by a single government letter — with zero advance notice and no contractual recourse. Standard cloud-API agreements and force-majeure clauses offer essentially no protection against a regulatory suspension of this kind, because almost none were drafted with this scenario in mind.

Four Steps to Reduce Your Exposure