AI Strategy• 10 min read

June 2026: The 18 Days Washington Switched Off Anthropic's Most Capable Models

Anthropic released Claude Fable 5 and Claude Mythos 5 on June 9, and three days later a Commerce Department export control directive forced it to switch both off for every customer in the world. The models came back 18 days later, after protests from security researchers and allies, new competition from Asian rivals, and fresh security commitments from Anthropic. June showed that access to frontier AI is now a policy variable, not just a product decision.

Author: Macaulan Chiaramonte

June 2026 should have been remembered for a launch. On June 9, Anthropic released Claude Fable 5, which it called "a Mythos-class model that we've made safe for general use," alongside Claude Mythos 5, the same underlying model with some safeguards lifted for vetted cyber defenders.

Instead, June became the month the U.S. government recalled a commercial AI model. At 5:21pm ET on June 12, Anthropic received a directive barring access by any foreign national, and because it could not check nationality in real time, it disabled both models for all customers. Law firm Greenberg Traurig called it "the first known U.S. use of export control authorities to regulate a particular AI frontier model." The controls were lifted on June 30.

Fable 5 and Mythos 5: One Model, Two Products

Anthropic's launch post said Fable 5's capabilities "exceed those of any model we've ever made generally available." The company said it scored highest among frontier models on Cognition's FrontierCode evaluation and set the top score on Hebbia's Finance Benchmark. Both models were priced at $10 per million input tokens and $50 per million output tokens, "less than half the price of Claude Mythos Preview."

What made the launch unusual was its structure. Fable 5 carries safety classifiers that watch for requests involving cybersecurity, biology, chemistry and model distillation, and when they trigger, "the response is automatically handled by Claude Opus 4.8 instead." Anthropic said the classifiers fire "on average, in less than 5% of sessions."

Mythos 5 ships without those classifiers. Anthropic said it has "the strongest cybersecurity capabilities of any model in the world" and deployed it through Project Glasswing, its restricted program for cyber defenders (launched in April, when Anthropic held back Mythos Preview), "in collaboration with the US government." In effect, the company split one model into a public product and a controlled one, separated by safeguards rather than by capability.

The Backdrop: A Voluntary Framework Signed June 2

A week before the launch, President Trump signed Executive Order 14409, "Promoting Advanced Artificial Intelligence Innovation and Security." According to the White House fact sheet, it creates an AI cybersecurity clearinghouse to find and fix software vulnerabilities at scale, and a classified benchmarking process to identify "covered frontier models" with advanced cyber capabilities.

For those models, the order directs the government to build "a voluntary framework in collaboration with AI developers" that gives trusted partners secure early access. The fact sheet is explicit that nothing in the order authorizes "any mandatory governmental licensing, pre-clearance, or permitting requirement for the development, publication, release or distribution of AI models."

That language is the key to reading what happened next. The order described a cooperative, opt-in relationship with frontier labs. Ten days later, a different instrument, export control law administered by the Commerce Department, produced something that worked very much like a license requirement on a model that was already in customers' hands.

June 12, 5:21pm: The Directive That Took Fable 5 Offline

Anthropic's statement that evening said the government had directed it to "suspend all access to Fable 5 and Mythos 5 by any foreign national, whether inside or outside the United States," including Anthropic's own foreign national employees. According to CSIS, the letter made access conditional on "an approved export license from the Bureau of Industry and Security (BIS)."

No global AI service can sort its users by citizenship overnight. As Anthropic later explained, "the order took effect immediately and we had no reliable way to verify nationality in real-time," so it suspended both models for every user, inside the United States and out.

The company complied but disagreed in public. Its statement noted that "the letter did not provide specific details of its national security concern," and said: "we disagree that the finding of a narrow potential jailbreak should be cause for recalling a commercial model deployed to hundreds of millions of people."

"Fix This Code": The Jailbreak Behind the Order

The trigger was a report from Amazon. In its June 30 account, Anthropic said Amazon researchers had found a way around Fable 5's safeguards by prompting it to identify software vulnerabilities, and that "in one case, the model produced code demonstrating how the relevant vulnerability could be exploited."

Fortune reported the specifics. When the researchers asked Fable to "review the code for security issues," the model refused; when they asked it to "fix this code," it produced patches. Fortune also reported that the finding reached the administration "including in a phone call Amazon CEO Andy Jassy had with the White House."

Katie Moussouris, founder of Luta Security, reviewed Amazon's findings at Anthropic's request. She wrote that the issue "cannot meaningfully be fixed, and any attempt would only weaken the model for defense," because "defenders need to be able to ask AI to fix bugs in a file, explain why the fix matters, and write tests that confirm the patch works."

This is the dual-use problem in its plainest form. A model that can fix a vulnerability has, by definition, found it, and an explanation of the fix is a short step from an explanation of the exploit. Any policy that treats vulnerability reasoning as the controlled capability will collide with the everyday work of secure software development.

Export Rules Built for Shipments, Applied to an API

The legal basis was contested from the start. CSIS analysts reported that the letter cited Section 744.22 of the Export Administration Regulations (EAR), which allows "is informed" controls where there is a risk of military-intelligence end use, along with Section 734.13, which defines what counts as an export.

Their central objection was that "it is unclear how user access to models equates to a release of EAR software or technology." They also pointed out that Commerce had previously issued advisory opinions treating remote access transactions as not subject to the EAR at all.

The practical warning in the CSIS analysis is the one enterprises should take most seriously: "every company must consider the possibility of such controls being used again." Because the directive arrived as a letter to one company rather than as a published rule, it offers no public thresholds that other labs, or their customers, can plan around.

Defenders and Allies Push Back

The security community reacted first. An open letter signed by veterans including Alex Stamos, Casey Ellis, Paul Vixie, Rachel Tobac and Moussouris had 76 signatures when TechCrunch reported on it on June 15. It argued that "this action has taken the best models away from [cybersecurity] defenders," and that doing so "without a good reason when our adversaries are rapidly advancing is dangerous."

Allies followed. At the G7 summit in Évian-les-Bains on June 17, Dario Amodei urged democracies to "resist the temptation to splinter," Sam Altman said G7 nations should have access to AI-powered cybersecurity tools, and Demis Hassabis also called for collaboration, according to Fortune. French President Emmanuel Macron said the decision had "clarified the stakes," because one government could "turn off the switch."

In Britain, MP Al Carns was blunter, as quoted by IAPP: "This week the most advanced AI model on the planet got switched off by a foreign government." The same IAPP analysis noted that the European Commission had proposed a European Technological Sovereignty Package covering semiconductors, AI, cloud and open source. Dependence on U.S. models is now being framed as a sovereignty question.

Rivals Move Into the Gap

Competitors noticed too. Tokyo-based Sakana AI launched Fugu, which it says "stands shoulder-to-shoulder with leading models like Anthropic's Fable 5 and Mythos Preview," and whose website advertised "delivering frontier capability without the risk of export controls." A Sakana spokesperson told TechCrunch the timing was "entirely coincidental."

The same week, Chinese cybersecurity firm 360 reportedly unveiled Tulongfeng, a vulnerability-discovery tool it says can go head-to-head with Mythos. Citing Reuters, TechCrunch reported that 360 founder Zhou Hongyi described vulnerability-finding AI as a national strategic asset.

The commercial stakes were large. TechCrunch noted that Anthropic's run-rate revenue crossed $47 billion in May 2026, and every week the models stayed dark gave international customers a reason to test alternatives. By June 30, TechCrunch reported, the government was "under pressure to ease its restrictions on Anthropic to ensure that American AI could compete globally."

The Reversal: Commitments Instead of a Rule

The unwinding came in two steps. On June 26, the government approved restoring Mythos 5 for a set of U.S. organizations, and on June 30, Commerce lifted the export controls on both models. Anthropic said it would begin restoring Fable 5 globally on July 1 across the Claude Platform, Claude.ai, Claude Code and Claude Cowork.

Commerce Secretary Howard Lutnick said Anthropic "has agreed to proactively detect and address security risks associated with the models; to work diligently with the U.S. government on protocols and standards and releases for Mythos, Fable and future models; and to inform the US government of any malicious activity."

Anthropic's own account lists four commitments: expanded early access for designated government partners, rapid sharing of jailbreak and misuse information, scaled-up joint research, and shared security evaluation standards across the industry. It said a retrained classifier now blocks the Amazon technique "in over 99% of cases," and that researchers from Commerce's Center for AI Standards and Innovation (CAISI) tested both the old and new safeguards.

Anthropic also proposed a consensus industry framework for judging jailbreaks against four criteria: capability gain, breadth of capability gain, ease of weaponization and discoverability. If adopted, it would give labs and governments a shared vocabulary for deciding when a bypass justifies pulling a model, which is exactly what was missing on June 12.

What It Means for Enterprise and Technology Leaders

For organizations building on frontier models, June turned a theoretical risk into an operational one. Baker McKenzie partner Brian Hengesbaugh told IAPP that "companies should start to incorporate potential unavailability of AI models into company workflows." Five practical lessons follow:

  • Plan for unavailability: Treat a frontier model like any critical dependency, with a documented fallback and a tested switchover.
  • Engineer the fallback path: Fable 5 already hands some requests to Opus 4.8. Apply the same pattern across vendors, with evaluations that show what degrades.
  • Map who touches the model: Greenberg Traurig flagged API integrations, foreign national employee access and indirect access through automated workflows.
  • Know your security tooling: Vulnerability discovery is the capability governments now watch most closely, so inventory the AI tools your security teams rely on.
  • Track commitments, not just rules: Future access may hinge on pre-release government evaluation agreements as much as on published regulation.

None of this argues against adopting the most capable models. It argues for adopting them the way enterprises adopt any concentrated supplier: with redundancy designed in from the start, rather than added after the first outage.

Also in June

  • Apple introduces Siri AI (June 8): At WWDC, Apple unveiled "an entirely new version of Siri" with personal context and onscreen awareness, due as a beta later this year.
  • Five Eyes warn on AI cyber risk (June 22): Six cyber agencies from the U.S., UK, Canada, Australia and New Zealand warned that "the timeline is not years, it is months."
  • OpenAI's first custom chip (June 24): OpenAI and Broadcom unveiled Jalapeño, an inference processor OpenAI says shows better performance-per-watt in early testing.

Looking Forward: Frontier Access as Conditional Infrastructure

June 2026 used export control law against a single deployed model, then rolled it back in 18 days in exchange for commitments. That sequence extends the themes this series tracked in January, when agents began acting as autonomous workers and AI infrastructure became a strategic asset. Once a model is infrastructure, its availability becomes a matter of state.

Key developments to watch in the coming months:

  • A rule, or more letters: Whether Commerce publishes criteria for when a model triggers controls, or keeps acting case by case.
  • The jailbreak framework: Whether other labs and CAISI adopt Anthropic's four severity criteria as a shared standard.
  • EO 14409 in practice: How the voluntary framework for pre-release access to covered frontier models takes shape.
  • Sovereign AI demand: Whether European and Asian buyers shift spending to local models after watching a U.S. model go dark.
  • Mythos-class tools elsewhere: Tools like Tulongfeng suggest vulnerability-finding AI will not stay a U.S. monopoly for long.

The 18 days in June did not end with a new rulebook. They ended with a negotiated settlement between one lab and one government, and a precedent that every company building on frontier AI now has to price in. The lesson for technology leaders is not that the next shutdown will happen, but that it can, and that resilience against it is now part of AI architecture.

Sources

Related Articles

AI Strategy

March 2026: Anthropic Takes the Pentagon to Court Over Who Sets AI's Limits

The Pentagon formally designated Anthropic a supply chain risk, a label never before applied to a domestic company, and Anthropic sued. On March 26, a federal judge blocked the designation and the federal ban on Claude, calling the government's actions "classic illegal First Amendment retaliation."

9 min read
AI Strategy

February 2026: AI Agents Trigger the SaaSpocalypse and a Software Reckoning

Anthropic's Cowork plugins, OpenAI's Frontier platform and Claude Code Security set off a month of steep selling in software, data and cybersecurity stocks. The rout was dubbed the SaaSpocalypse: a bet that AI agents will erode the per-seat licenses the software industry was built on.

10 min read