The Lutnick Letter Is Public Now. The Jailbreak Was Never the Real Reason.

Three days ago, when we covered the directive that pulled Fable 5 and Mythos 5 off the market, the central mystery was which jailbreak Commerce relied on — the narrow code-review technique Anthropic described, or the much louder Pliny the Liberator bypass sitting on GitHub. That framing assumed the jailbreak was the reason.

Now the letter is public. Bloomberg obtained the full text on June 16, and it changes the question. The jailbreak was the trigger. It was never the reason.

What the letter actually argues

The document is not a product-safety recall. It is an export-control order, and its load-bearing legal idea is a doctrine called deemed export.

Under deemed-export rules — the same regime that governs which foreign engineers can stand in a room with a controlled semiconductor design — handing a foreign national access to controlled technology on US soil is legally identical to shipping that technology to their home country. A Chinese national querying Mythos 5 from a desk in Palo Alto is, in this reading, an export to China. That is why the order’s scope is citizenship, not geography. It is also why Anthropic could not partially comply: there is no border to fence when the “export” is a logged-in API call.

The letter, per Bloomberg and the secondary legal reporting, warns of both civil and criminal penalties if Anthropic allows nationals of countries of concern to reach the model weights or the training compute. That is a materially heavier instrument than the public knew on June 13, when all anyone had was Anthropic’s own statement that it had received a “directive.”

Read against the letter, Anthropic’s statement now looks like it was answering a question the government wasn’t really asking. Anthropic argued the jailbreak was too narrow to justify a recall. The letter’s logic doesn’t actually hinge on the jailbreak’s severity. It hinges on the premise that a frontier model is controlled military-relevant technology, and that the jailbreak merely proved the models can be steered toward weapons-relevant output — vulnerability discovery, exploit generation, chemical synthesis routes. Once you accept that premise, the breadth of any single jailbreak is almost beside the point. The capability is the hazard; the jailbreak is the demonstration.

Why this is a bigger deal than a recall

The June 13 story was “the government pulled a model.” The June 16 story is “the government has a theory for pulling a model, and the theory generalizes.”

If a frontier model is controlled technology subject to deemed-export rules, then every US AI lab now faces a standing obligation that has nothing to do with whether their model has a known jailbreak today:

• The customer base is the compliance surface. Any lab serving foreign nationals — which is all of them — is, under this doctrine, continuously exporting. The question stops being “is our model safe” and becomes “can we prove who every user is.”
• Open weights become legally radioactive. A downloadable open-weight frontier model cannot enforce a citizenship gate at all. Meta’s Llama line, Mistral’s open releases, and any US lab shipping weights are the most exposed to this framing, not the least.
• The penalty is criminal, not just commercial. A civil fine is a cost of doing business. Criminal exposure for executives changes how fast a company complies and how hard it argues. Anthropic disabled both models within hours — that speed reads differently once you know criminal liability was on the table.

The resolution path, and what it reveals

The negotiation is the tell. Senior Anthropic technical staff met Commerce officials in Washington on Monday, June 15 — and the working-level meeting was joined by National Cyber Director Sean Cairncross. That is not who shows up for a product dispute. The administration’s reported position is that the models stay locked until the US government’s own national-security apparatus is “hardened,” something officials expect to take a few weeks.

Parse that. The government is not saying “Anthropic must fix the jailbreak.” It is saying the government itself needs time to prepare before these models can safely exist in the market again. The bottleneck being described is state readiness, not vendor patching. That is consistent with the deemed-export reading and inconsistent with the jailbreak-recall reading. You don’t harden a federal cyber posture to wait out a bug fix.

What changed since June 13

Three things we did not know then, and know now:

One: it came from the top. Not a faceless agency directive — a letter signed by the Commerce Secretary to the CEO by name. This is being run as policy, with a principal’s signature on it.

Two: the legal hook is deemed export. The order restricts by citizenship because, under this doctrine, the user is the export. That is the detail that makes the global shutdown the only compliant move, and it is the detail with the longest reach into the rest of the industry.

Three: the stakes are criminal. The earlier reporting implied regulatory pressure. The letter reportedly names criminal penalties. That reframes Anthropic’s instant compliance from “cautious” to “rational.”