Commerce Export Order Took Claude Fable and Mythos Offline

On the evening of June 12, at 5:21 p.m. Eastern, the U.S. government sent Anthropic a letter. By midnight, the company's two most powerful models — Claude Fable 5 and Mythos 5 — were gone for everyone. Not just users in Beijing. Not just foreign contractors. Everyone. A directive nominally aimed at foreign nationals had, in practice, flipped off frontier AI for the entire planet because Anthropic runs across dozens of cloud environments — AWS Bedrock, Google Cloud, Microsoft Foundry, Snowflake, Box, its own API — and has no real-time mechanism to verify citizenship at the point of inference. Compliance meant universal shutdown. Three days after launch.

This is the moment frontier AI stopped behaving like a product release and started behaving like geopolitical infrastructure. Anthropic had unveiled Fable 5 on June 9 as its first public Mythos-class model — a consumer-facing version of an architecture the company had previously considered too capable in cybersecurity to release broadly. Mythos 5, the less restricted variant, was available only to vetted partners through Project Glasswing. Both were offline by Friday night. Opus, Sonnet, and Haiku kept running. The kill switch, as French President Emmanuel Macron called it at the G7 summit this week, had been pulled on the crown jewels.

The stated trigger is a jailbreak — or, more precisely, a dispute about what counts as one. According to Axios and Forbes, Commerce Secretary Howard Lutnick ordered Anthropic to suspend access for any foreign national anywhere in the world, including foreign nationals employed by Anthropic inside the United States. The government's concern, per Anthropic's public statement, centers on a technique for bypassing Fable's safety guardrails — essentially asking the model to read a codebase and fix software flaws. Anthropic reviewed the underlying report and argued the capability displayed is "widely available from other models (including OpenAI's GPT-5.5), and is used every day by the defenders who keep systems safe." White House AI adviser David Sacks told a different story on X: "A highly credible, trusted partner of both Anthropic and the USG who was testing Fable came forward with a jailbreak of those guardrails. The Admin asked Dario to fix the jailbreak or de-deploy the model. Dario refused." Anthropic disputes both the severity and the characterization of its response. The letter itself offered no public technical detail. The models went dark anyway.

If the jailbreak is the proximate cause, the backdrop is a months-long feud that made Anthropic an obvious first target. In February, the Pentagon designated Anthropic a supply-chain risk — the first American company to receive a label historically reserved for foreign adversaries — after the company refused to allow Claude for mass domestic surveillance or fully autonomous lethal weapons. A federal judge in California blocked much of that designation in March, calling it likely retaliatory; a separate challenge in the D.C. Circuit remains unresolved. Defense Secretary Pete Hegseth posted on X this week that "every passing day proves" the Pentagon was right to eject Anthropic. Sacks insisted the export order is unrelated. The industry is not buying the separation. When the same administration can blacklist a lab in February and switch off its flagship models in June, the distinction between procurement punishment and export control starts to look academic.

The Amazon angle makes the story stranger still. Forbes, Semafor, and the Wall Street Journal report that Amazon CEO Andy Jassy flagged security concerns to Treasury Secretary Scott Bessent and other officials after Amazon researchers tested Fable 5 — Amazon being Anthropic's largest investor and primary cloud host. A separate reporting thread suggests the White House also acted over fears that Mythos access had reached a South Korean telecom partner with alleged ties to China. Anthropic says it revoked that partner's access when flagged. Either way, the enforcement mechanism is the same: an export-control letter, not a product recall process, not a transparent safety review, not the statutory framework CEO Dario Amodei had publicly advocated just days earlier. In a policy essay published June 10 — two days before the shutdown — Amodei argued the government should have authority to block unsafe frontier deployments through a "transparent, fair, clear" process "grounded in technical facts." He opened with a Lord of the Rings analogy about Treebeard, the sentient tree who moves too slowly for the Hobbits' urgency. This week, Treebeard moved fast.

The global reaction was immediate and widespread. At the G7 in Evian, India's Narendra Modi called for "broad and inclusive" access to American AI models; the UK reportedly sought a carve-out and was rejected. Macron warned allies would not buy models that can be switched off overnight. German ifo Institute president Clemens Fuest noted Europe controls less than 5% of global AI infrastructure against roughly 75% for the United States. More than 170 tech executives signed an open letter warning the curbs risk America's AI leadership; separately, 76 cybersecurity veterans — including Alex Stamos, Paul Vixie, Katie Moussouris, and Sophos CEO Joe Levy — told Lutnick and National Cyber Director Sean Cairncross that pulling the best tools from defenders while adversaries advance is "dangerous." CSIS analysts Kate Koren, Kevin Kurland, and Aalok Mehta went further on the legal question: Commerce may have used export authorities that have "never been used before," on statutory grounds that lack an implementing regulation, to impose a worldwide control that the Export Administration Regulations may not clearly authorize for remote model access. The action, they wrote, "is likely to broadly undermine goals for global adoption of U.S. models."

For markets, the timing could hardly be worse. Anthropic filed a confidential IPO registration earlier this month, with reporting putting revenue near $47 billion and valuation approaching a trillion dollars. Bankers have floated a $30–$60 billion offering that would rank among the largest listings in history — in the same season OpenAI submitted its own S-1 and SpaceX raised $75 billion in a frenzy of AI-adjacent public appetite. IPO expert Jay Ritter told Deutsche Welle that Kalshi markets still priced an 85% chance of an Anthropic IPO before November 1, suggesting investors expect a resolution. That confidence may be rational. Negotiations between Anthropic and the White House have continued since the shutdown; the company's international managing director said at a June 18 Seoul office opening that he was "very confident" both models would return "in the coming days." Reporting points toward a July 8 identity-verification rollout that could restore access for verified U.S. persons. But even a partial restoration leaves the precedent intact: Washington demonstrated it can halt a commercially deployed frontier model mid-flight, on the basis of a security assessment the company disputes, through legal authorities experts question, with almost no public evidence.

Convina's view: This is not primarily a story about whether Anthropic's guardrails were good enough. It is a story about what happens when the world's most capable software has no mature off-switch except national security law written for a different era. The administration spent its first year rolling back AI guardrails to compete with China — then improvised its first enforcement action against the lab that marketed safety most aggressively. That inconsistency is the policy. Enterprises that treated Claude, GPT, or Gemini as interchangeable API endpoints just learned they are renting access to assets the U.S. government can reclassify as export-controlled munitions on a Friday evening. Boards should add a new line to vendor risk: not just uptime and data residency, but sovereign discontinuation risk — the probability that a model you built a workflow on can be switched off by a letter you never see. The kill switch was always theoretical. Anthropic proved it works. The question now is whether anyone building on American frontier AI believes the switch will stay off.