21SEP

Washington pulls a live AI model

4 min read

16:45UTC

At 5:21pm Eastern on Friday 12 June, Commerce Secretary Howard Lutnick ordered Anthropic to bar foreign nationals from Claude Fable 5 and Mythos 5. The only way to comply was to switch both off for everyone on earth.

← Back to AI closes the hiring door, not the exit Jump to analysis ↓

EconomicDeveloping

Key takeaway

Washington used an export law built for chips to switch off a consumer AI model hundreds of millions were using.

At 5:21pm Eastern on Friday 12 June, Commerce Secretary Howard Lutnick sent Anthropic chief executive Dario Amodei a directive barring every foreign national, inside or outside the United States, from accessing Claude Fable 5 and Claude Mythos 5 ¹. The order reached Anthropic's own foreign-national staff. To comply, the company disabled both models for every customer worldwide, three days after Fable 5 went on sale.

Anthropic is the San Francisco AI lab behind the Claude family of models; Commerce is the US federal department that polices what technology may leave American hands. Lutnick reached them through the deemed-export doctrine: under US law, granting a foreign national access to controlled technology on American soil counts as an export to that person's home country ². A live consumer model cannot check the nationality of every user mid-conversation, so the only way Anthropic could keep foreign nationals out was to keep everyone out. The controlled item, in effect, became the act of using the model rather than possessing it, and no commercial AI product is built to gate inference by passport in real time.

The capability did not change on Friday. This desk has tracked the Mythos arc since April, when the Treasury and the Federal Reserve summoned five Wall Street bank chiefs over the withheld Mythos Preview, the model Anthropic distributed to twelve partners under Project Glasswing, that the UK AI Security Institute clocked at a 32-step autonomous attack chain, and that Anthropic then reframed in a 244-page Alignment Risk Update abandoning its own capability thresholds. By 1 May, OpenAI's rival GPT-5.5 had cleared the identical 32-step chain , and it remains on sale. The trigger, per Axios, was an unnamed rival claiming to have jailbroken Mythos 5 by asking it to read a codebase and fix the flaws it found; the administration first sought a launch pause, Anthropic refused, and the letter followed ³.

The campaign-money symmetry sharpens the irony. Anthropic backed the pro-regulation Public First campaign while the OpenAI camp funded the anti-regulation Leading the Future PAC . The lab that asked to be regulated is the one Washington reached first, while the untouched competitor is the one whose founders paid to weaken the rules. Restoration terms, whenever they come, will set the first precedent for how a deployed consumer model re-enters a market after a national-security recall.

Deep Analysis

In plain English

The US government has a rule called the deemed-export doctrine. Under it, if an American company lets a foreign citizen use controlled technology inside the United States, the law treats that as if the technology had been physically shipped to that person's home country. The same rule applies to military hardware, advanced chips, and now, apparently, AI models. The problem is that Claude cannot tell who is foreign and who is not when someone types a question. It has no passport scanner. So when Commerce Secretary Howard Lutnick sent Anthropic a letter saying no foreign nationals could use Fable 5 or Mythos 5, the only way Anthropic could comply was to switch both models off for everyone, everywhere, including paying customers in every country. The government did not seize a server or a file. It made access illegal, and the model's inability to enforce that distinction did the rest.

Deep Analysis

Root Causes

The deemed-export doctrine was written for an era when controlled technology lived in a discrete artifact: a chip, a weapon schematic, a software binary on physical media. The 2025 AI Diffusion Rule extended it to closed model weights, which at least travel as discrete files.

Runtime inference on a cloud API produces no artifact that crosses a border; the foreign-national user receives a token stream, not a weight file, meaning the 'export' the law notionally controls exists only as a probabilistic output. No commercial inference architecture gates by nationality in real time because no commercial architecture was designed to be a border control.

A second structural vulnerability sits in the competitive reporting channel. An unnamed rival claimed the Mythos jailbreak and routed the report to the administration.

Private competitor reporting to regulators creates a repeatable incentive: flagging a rival's vulnerability is costless to the reporter, imposes a compliance burden on the target, and carries an asymmetric political return if the target lobbied for regulation. The BIS directive reached Anthropic, not OpenAI, despite GPT-5.5 carrying the same vulnerability at the same date.

Escalation

Escalatory. The directive represents the first instance of a US government withdrawing runtime access to a deployed consumer AI product, not a pre-release model or a hardware export. Restoration terms, if and when set, will define whether the control persists, widens, or is abandoned as unenforceable, each of which has a different escalatory profile for the broader AI sector.

What could happen next?

Precedent
Lutnick's directive establishes that the deemed-export framework can reach runtime inference access, not only model weights or hardware, giving the BIS legal surface to act on any deployed frontier model used by foreign nationals.
Immediate · Assessed
Risk
Competitors may use the private-reporting channel to the BIS as a competitive lever, filing jailbreak claims against rivals to trigger suspension events, given that Anthropic's suspension occurred while GPT-5.5 carried the same vulnerability.
Short term · Suggested
Consequence
Enterprise procurement of US-based frontier AI models now carries a regulatory-withdrawal risk that did not exist before 12 June, which will raise the cost and complexity of multi-year AI infrastructure commitments.
Medium term · Assessed
Opportunity
Non-US frontier labs, particularly those in the EU, UK, and Canada, can now market regulatory-jurisdiction diversity as a feature to international enterprise customers who cannot afford a repeat suspension.
Medium term · Suggested

Source Landscape

This story draws on mixed-leaning sources from United States

United States

Primary parallel: Between 1993 and 1999 the Clinton administration classified strong encryption software as a munition under the International Traffic in Arms Regulations, barring export to foreign nationals. By 1996 Netscape, RSA and PGP were all complying with export-limited builds.

The controls collapsed not because courts struck them down but because the same code was published overseas and available worldwide; the State Department dropped encryption from the munitions list in 2000, accepting that globally distributed software was technically uncontrollable.

Counter-parallel: Export controls on EUV lithography machines and advanced semiconductor manufacturing equipment have stuck. ASML ships from the Netherlands, not from American soil, yet US pressure via the Foreign Direct Product Rule has kept the machines out of Chinese fabs since 2023.

The difference: EUV machines are physical, capital-intensive, have roughly 50 global customers, and cannot be replicated by downloading a file. A language model running on globally distributed servers is structurally closer to 1990s encryption code than to a lithography machine.

Fable 5 was priced at $10 per million input tokens and $50 per million output tokens, with free access inside paid plans until 22 June.

Stripe's publicly demonstrated use case, a 50-million-line code migration completed in one day against a human team's two-month estimate, implies a task-input size that at $10 per million tokens would cost several hundred dollars and return weeks of developer labour. At scale, enterprise developers were beginning to price cloud inference as a substitute for junior engineering headcount.

A global suspension disrupts Anthropic's revenue and the procurement decisions of enterprises that had built Fable 5 into their workflow pipelines. Enterprise contracts with committed token volumes become voidable force-majeure claims. The precedent that regulatory action can withdraw a deployed consumer AI product with no notice period raises the risk premium on long-term AI infrastructure commitments across the sector, reaching beyond Anthropic to any US-jurisdictioned AI provider.

Consensus view: The Cato Institute's Inu Manak and Georgetown Law's Chris Mibrady both assessed the Lutnick directive as an expansive reading of the Bureau of Industry and Security's deemed-export authority, stretching a doctrine designed for hardware transfers and closed model weights into real-time inference access, a category the 2025 AI Diffusion Rule never addressed.

Counter-view: Tsinghua University's Institute for AI International Governance argues the action is consistent with US strategic logic: the purpose of the deemed-export framework has always been to prevent foreign nationals gaining capability, not merely to stop hardware crossing a border.

Beijing's own export controls on frontier chip designs follow identical reasoning, and Chinese scholars note the US is applying to AI the same controls it already applies to satellite and nuclear technology access on domestic soil.

Key tension: Whether the deemed-export doctrine, once applied to runtime inference, can be enforced at all on a globally deployed consumer product, or whether architectural reality forces any such control to become a total withdrawal by default.

First Reported In

Update #13 · Washington pulls a live AI model

Anthropic· 13 Jun 2026

Read original →

Causes and effects

Caused by

AI super PACs spend $150m on primaries

The Leading the Future vs Public First PAC split frames the political irony of the export directive: the lab that backed regulation is the first one hit.

Occurred 2 Jun 2026

Read story →

AISI: GPT-5.5 matches Mythos on 32-step attack

AISI confirmed GPT-5.5 cleared the same 32-step attack chain on 1 May, the capability comparison Anthropic cited in disputing the selective application of the directive.

Occurred 1 May 2026

Read story →

This Event

Washington pulls a live AI model

For the first time, a US government has reached past studying an AI model and used export law to turn off a consumer product hundreds of millions were already using.

Led to

Anthropic AI ban enters its second week

The Lutnick directive of 12 June caused the ongoing suspension that entered its second week, triggering House demands, G7 discussions, and Washington negotiations.

Occurred 18 Jun 2026

Read story →

Different Perspectives

Salesforce, Synopsys and TD Bank Group

Salesforce, Synopsys and TD Bank Group each filed quarterly disclosures in late August booking restructuring charges, or none at all, without naming AI as a cause. Their silence matters because Challenger's tracker shows AI as a stated reason fell to fourth place in August even as the year-to-date AI-cut total still leads at 116,175.

Singapore, South Korea, Taiwan and Indonesia

Singapore launched its Skills and Workforce Development Agency on 16 September, giving citizens six months of free premium AI tools, while South Korea ring-fenced its AI tax windfall in a new Future Response Fund. Taiwan kept funding its AI build past NT$190bn and Indonesia rewired vocational training around AI literacy, betting state-built skills beat a market-led adjustment.

ver.di, CGT Fonction Publique and CCOO

Germany's ver.di banked a 3.3% pay rise on 1 September and opened talks on a Tarifvertrag Transformation covering dismissal bans and reskilling, while France's CGT rejected Paris's AI negotiating timetable the same week. Spain's CCOO went further on 21 September, proposing to tax companies by the jobs they generate rather than wait for the next bargaining round.

BIS General Manager and Federal Reserve governors

The BIS's General Manager said on 10 September that AI displacement remains limited, even as the BIS's own survey found nearly 80% of firms plan to automate roles. Two Federal Reserve governors made the same point in July, arguing the labour-market data does not yet show a mass-firing event.

Bank of Canada, ONS and ECB

The Bank of Canada found the job-finding gap between AI-exposed and unexposed occupations widened from 2.2 to 13.9 percentage points since 2015-19, while separations barely moved. That framing, a hiring freeze rather than a firing wave, is echoed by the ECB's finding that euro-area AI use hit 52% of workers in 2026, concentrated among the university-educated.

Office for National Statistics

Deferred its Transformed Labour Force Survey beyond November 2027 and disclosed a May 2026 telephone-collection failure. The ONS carries no AI-attribution layer at all, so Britain sits outside this month's cohort of measuring states by its own admission.