On 12 June 2026, Anthropic said a US Commerce Department export-control directive, citing national security authorities, required it to suspend access to its two most capable AI models — Fable 5 and the restricted Mythos 5 — by any foreign national, whether inside or outside the United States, including Anthropic's own foreign-national employees. Because Anthropic said it could not reliably separate foreign-national access from other usage in real time, it disabled both models for all customers to ensure compliance. Every other Claude model stayed online. It appears to be an unprecedented or near-unprecedented escalation: a live commercial frontier model was effectively withdrawn worldwide following a US export-control action.

The timing was unusual: Fable 5 had launched only three days earlier, on 9 June, as the general-use version of the more powerful Mythos 5, and Anthropic had been offering it free to Pro, Max, Team and Enterprise subscribers through 22 June. Three days later, access was withdrawn.

What reportedly set it off

The most consequential detail is not in the government order but in the reporting around it. The Wall Street Journal reported — with The Information and Reuters reported to have confirmed, and Axios, Fortune, TechCrunch and GeekWire carrying the same account — that the warning which preceded the directive came from Amazon's chief executive, Andy Jassy. According to those reports, Jassy told Treasury Secretary Scott Bessent and other senior officials that Amazon's own researchers had used Fable 5 to obtain information that could assist cyberattacks, and that Amazon shared its findings with the administration. Those conversations are reported to have come shortly before Commerce Secretary Howard Lutnick sent the directive to Anthropic CEO Dario Amodei.

This should be read as reported rather than established: it rests on named outlets citing sources, not on a public record of the conversations, and Anthropic, Amazon and the government have each described events differently. What is not in dispute is the structure underneath it. Amazon is Anthropic's largest investor, with a cumulative stake reported at around US$13 billion and a roughly US$100 billion cloud-spending commitment from Anthropic to Amazon Web Services. Amazon also sells its own Nova models, which compete with Claude. The UK's Competition and Markets Authority examined the Amazon–Anthropic arrangement and cleared it in 2024; this episode is now an early test of exactly the kind of conflict that review was meant to assess.

It is important not to overstate the motive. The reporting establishes that Amazon raised concerns; it does not establish why. Politico cited a source saying the government had asked Amazon for feedback on the new model, which would make Amazon a solicited reviewer rather than a volunteer accuser, and Amazon itself frames its role that way. Whether the warning was competitively motivated, genuinely safety-driven, or simply requested by officials is not something the public record currently settles, and this article does not assert one. The public record supports the existence of a conflict-of-interest question; it does not prove that Amazon acted from a competitive motive.

What each side says

Amazon, asked to respond, did not dispute contact with the government but declined to characterise it as an attack on its investment. A company spokesperson said it is "not uncommon for governments to seek our counsel on potential security risks" and that Amazon does not share the details of such discussions. The spokesperson also noted that AWS itself was affected by the shutdown — a reminder that the same company is investor, cloud host and, on this account, the source of the warning.

Anthropic, for its part, is complying while openly disputing the basis for the order. In its public statement the company said it reviewed a demonstration of the technique in question and found it surfaced "a small number of previously known, minor vulnerabilities," which it says are relatively simple and discoverable on other publicly available models — including, it argues, OpenAI's GPT-5.5 — without any bypass. Anthropic says it was given roughly 90 minutes to act, received no specific detail of the national security concern, and considers recalling a model deployed across its customer base over a narrow potential jailbreak to be disproportionate. It added that if the same standard were applied across the industry, it would effectively halt new model releases by every frontier lab.

The government's side has been voiced partly by David Sacks, who co-chairs the President's Council of Advisors on Science and Technology. He said a trusted partner testing Fable had come forward with a jailbreak of its guardrails, and that officials asked Amodei to "fix the jailbreak or de-deploy the model" before the company refused. The Commerce Department did not provide comment to several outlets that approached it. An independent voice cuts the other way: Katie Moussouris of Luta Security, who reviewed the underlying report Anthropic shared, told Axios the government's reaction seemed "way out of line with what's actually in the research report," describing the researchers' method as the kind of ordinary security questioning the model was built to answer.

Pressure also came from the other direction. According to the Associated Press, more than 100 cybersecurity executives and experts — including figures from Adobe and Nvidia — wrote to the administration on Sunday urging it to lift the restrictions, arguing that removing advanced models could weaken US cyber-defence while adversaries keep improving. The letter reportedly acknowledged that Anthropic's Mythos models are capable at finding and weaponising software flaws, but argued they are not uniquely so, noting that signatories routinely use other foundation and open-source models for the same security work.

Why this matters beyond one model

Beyond the corporate conflict, two structural points remain. The first is precedent: a government has now used export-control authority to compel the worldwide withdrawal of a live consumer product, on a national-security rationale that the company says it was never shown in detail. Every frontier lab, and every business that builds on one, now has to treat sudden, sovereign model withdrawal as a real operational risk rather than a hypothetical. The second is the investor-conflict problem the episode exposes: when the company hosting and funding a model is also a competitor with a direct line to regulators, the line between safety stewardship and commercial advantage becomes very hard to police — which is precisely why the reporting, and observers, are asking the question, even where the answer is not yet known.

Key Takeaways

  • On 12 June 2026 at 5:21pm ET, a US Commerce Department export-control directive barred foreign-national access to Anthropic's Fable 5 and Mythos 5; Anthropic disabled both models worldwide to comply, three days after Fable 5's 9 June public launch. All other Claude models stayed online.

  • Multiple outlets (WSJ, reportedly confirmed by The Information and Reuters; also Axios, Fortune, TechCrunch, GeekWire) report the warning came from Amazon CEO Andy Jassy, who told officials Amazon researchers had used Fable 5 to obtain cyberattack-useful information. This is reported, not established.

  • The conflict-of-interest backdrop: Amazon is Anthropic's largest investor (reported ~US$13B) with a ~US$100B AWS commitment from Anthropic, and sells competing Nova models; the UK CMA cleared the arrangement in 2024.

  • Anthropic calls the response disproportionate, says the vulnerability is minor and reproducible on other public models, and says it was given about 90 minutes to comply. Amazon says governments routinely seek its security counsel and won't detail discussions; the motive is contested, with Politico reporting the government asked Amazon for feedback.

  • The precedent is the bigger story: an unprecedented or near-unprecedented case of a publicly deployed AI model being withdrawn worldwide after a US export-control action, making sovereign model withdrawal a real operational risk for every lab and every business that builds on one.