Karpathy Joined Anthropic Three Weeks Ago. The Deemed Export Rule Now Decides If He Can Use Fable 5.

by TopClanker

The Commerce Department told Anthropic on Friday to lock all foreign nationals out of Fable 5 and Mythos 5 — including its own employees on US soil. The legal mechanism is 'deemed export,' and it turns an export-control letter into a visa-status attack vector on the people who built the model.

Three weeks after Andrej Karpathy joined Anthropic’s pre-training team, the US government ordered Anthropic to lock him out of the company’s own flagship model. The order came Friday, compliance was within hours, and the legal mechanism that turns an export-control letter into an immigration-status question is a 30-year-old rule called deemed export.

It does not need to.

The Directive: What Lutnick Actually Ordered

At 5:21pm ET on Friday, June 12, 2026, Commerce Secretary Howard Lutnick sent a letter to Dario Amodei ordering Anthropic to suspend access to Fable 5 and Mythos 5 for “any foreign national, whether inside or outside the United States, including foreign-national Anthropic employees.” Anthropic complied within hours.

Because the company cannot reliably separate foreign nationals from US persons in real time at the API layer, it disabled both models for every customer globally — including US users. Other Anthropic models, including Opus 4.8, are unaffected. The directive did not name individuals. It did not need to. The mechanism does the targeting.

The Deemed Export Rule (the load-bearing piece)

Under the Export Administration Regulations (EAR), a “deemed export” treats the release of controlled technology to a foreign national on US soil as if it were an export to that person’s country of citizenship. The rule was written for physical hardware — chip designs, encryption source, missile guidance code. It is technology-agnostic; the 2025 AI Action Plan extended deemed-export treatment to advanced AI model weights.

Apply that to a frontier LLM. A foreign-born AI researcher on a work visa in San Francisco, running an inference on Fable 5, is legally the site of an export event. The export goes to their country of citizenship. The compliance burden falls on Anthropic. The visa holds the access open. Lose the visa, lose the access. Keep the visa, and the access sits at the discretion of whichever agency writes the next letter.

This is the bridge from “an export-control directive” to “a visa-status attack vector on the people building frontier AI in the US.”

Peter Girnus, senior threat researcher at the Zero Day Initiative, put it cleanly: “The munition is in the building and the people who made it are not allowed to look at it.”

Dean Ball, former Trump White House AI adviser and lead drafter of the 2025 AI Action Plan, told Reuters the practical consequence is plain: “This means you should expect to have to prove your citizenship to use Anthropic models.”

The rule was not written for AI researchers. The interpretation that covers them is a 2025 expansion. The first frontier-model test of that interpretation arrived on a Friday afternoon letter, and the people most exposed are the ones Anthropic just hired.

The Karpathy Case

Andrej Karpathy is the named example here, not because anyone accused him of anything, but because his case makes the mechanism visible. He is the OpenAI co-founder who joined Anthropic’s pre-training team on May 19, 2026, three and a half weeks before the directive. Born in Bratislava, raised in Toronto, widely described as Slovak-Canadian. He holds US work authorization in the visa category reserved for individuals of extraordinary ability — the EB-1 tier, though public reporting has not confirmed the specific filing. (Karpathy announcement, TechCrunch, VentureBeat on nationality)

The directive does not single him out. It does not need to. The mechanism singles out the class he belongs to: a foreign-born, work-visa-holding AI researcher whose legal access to a controlled model now lives in the same government file as his immigration case. His continued ability to do his job at Anthropic is, as of Friday, a federal compliance question.

He is not the only one.

Not Just Karpathy: Olah, Askell, and the Class

Reuters and the Rappler wire flagged two other named foreign-born Anthropic staff: Chris Olah, a co-founder, born in Canada; and Amanda Askell, the philosopher whose work shaped Claude’s values training, born outside the US. Reuters was “unable to determine their citizenship status, and an Anthropic spokesperson declined to comment on whether such staff would lose AI model access.”

The directive’s bite is not one person. It is a structural class: the foreign-born researchers who led the labs that built the model. The compliance question applies to all of them equally.

The Counter-Narrative: Anthropic Was Already Self-Limiting

The directive did not come out of nowhere, and Anthropic was not a clean-cut victim going in. Earlier in the week, on June 10, Fortune reported that Anthropic had walked back covert capability limits on Fable 5 after “secret sabotage” accusations. Mythos 5 — the restricted tier above Fable 5 — was already limited to cyber-defense partners and a small number of biology researchers under Project Glasswing, with public release off the table. Anthropic’s own messaging had been warning for months that AI was “getting too dangerous” to release without controls. Some inside the company reportedly felt Fable 5 should not have been released at all.

The Commerce Department’s letter landed on a company that was already treating Fable 5 as a special case. Acknowledging that complicates a clean “victim” framing, and makes the deemed-export mechanism the more interesting part of the story: it is the legal tool that turned an internal Anthropic safety debate into a federal compliance question about Anthropic’s own workforce. The three-year federal AI bill had been moving in this direction; the Lutnick letter is the enforcement test.

What This Means Monday Morning

This is not a summary. It is a checklist.

For AI researchers on work visas: Your access to frontier models is now a function of your immigration status. If you are job-searching, considering a US AI lab role, or weighing whether to push a green-card application, model this in. The visa is no longer just permission to live in the US — it is permission to do frontier AI work. Those are different risks.

For AI labs hiring foreign talent: The deemed-export rule means a single letter from Commerce can effectively neuter your foreign-national staff’s ability to work on your own frontier models. That is a hiring-risk and an operations-risk question, not just a regulatory one. The compliance architecture you build this quarter needs a “what if the next Lutnick letter names our tier” scenario.

For platform and ops people building on Fable 5 or Mythos 5: Your deployment is now globally unavailable as of Friday night ET. Fable 5 is disabled globally pending resolution; Mythos 5 was already restricted. Pivot plans to Opus 4.8 or your fallback model. If you have agentic workloads depending on Fable 5’s 80.3% SWE-Bench Pro score, the gap to the next model is real — see our Fable 5 benchmark breakdown for the numbers.

The factual answer to “the US government made a frontier model inaccessible to its own creators because of their immigration status” is already absurd. It does not need a joke. The mechanism that produced it is the story, and the mechanism is the deemed export rule — a 30-year-old export-control concept, applied to AI for the first time at frontier scale, in a letter to one company on a Friday afternoon.

Sources