Anthropic has disabled Fable 5 and Mythos 5, its two most capable AI models, for every customer, foreign and American alike, after a US government export control directive tied to cyberattack research; the company is complying but says the flagged issues are 'relatively basic'…
Anthropic has disabled Fable 5 and Mythos 5, its two most capable AI models, for every customer, foreign and American alike, after a US government export-control directive ordered the company to cut off any foreign national from accessing the systems, including the company's own foreign-born staff. The order puts the question of who decides when a frontier AI model is too dangerous to deploy squarely in the hands of a single US agency, with no public rule yet for how that decision gets made.
The directive arrived at 5:21pm ET on June 15, 2026, according to Anthropic's own statement. It told Anthropic to suspend Fable 5 and Mythos 5 for any foreign national, a category broad enough to sweep in Anthropic's own engineers working outside the United States. Anthropic is complying, but only by disabling the two models for all customers, since it has no workable way to give American users access while keeping foreign nationals out. Other Anthropic models are not affected.
What triggered the order, according to a Wall Street Journal report summarized by the TLDR AI newsletter, was a chain of conversations between Amazon CEO Andy Jassy and Trump administration officials. Amazon researchers had reportedly prompted Fable 5 with questions designed to surface information useful for cyberattacks, and Amazon escalated the results to Washington. The Wall Street Journal piece itself is paywalled, and the framing that Amazon's talks with the White House triggered the directive rests on TLDR's digest of that single article. Anthropic's public statement does not mention Amazon by name. Until the underlying article can be read in full, the Amazon-to-White-House chain should be treated as a reported account, not an established cause.
Anthropic's public response is sharp. The company says the capabilities at issue are "relatively basic" and that the same vulnerabilities can be triggered on other publicly available frontier models, including OpenAI's GPT-5.5. Anthropic points to its own defense-in-depth posture on Fable 5 and Mythos 5: 30-day data retention by default, narrow non-universal jailbreaks that do not transfer across topics, monitoring and shutdown capability, and thousands of hours of pre-launch red-teaming with US government, UK AI Safety Institute, and third-party evaluators. On that basis, the company warns that the government's standard, applied industry-wide, "would essentially halt all new model deployments for all frontier model providers."
The narrower news is that two named models went dark. The wider news is the standard the directive implies: if a model can be steered toward cyberattack help, the federal government now treats that as a controlled capability, and the only major lab willing to put a detailed objection on the record says the rule is unworkable at frontier scale. That is the precedent the rest of the industry has not yet had to answer to.
The mechanics matter. An export-control directive is a national-security tool traditionally used for hardware, dual-use chemicals, and weapons components, not for software customers log into. Applying it to a deployed consumer model, especially one that can also be reached from inside the United States, forces a clumsy choice: take the model down for everyone, or build a foreign-national firewall that has no real precedent in the consumer AI market. Anthropic chose the first option, and acknowledged in the same statement that the second was infeasible.
What to watch is the standard, not just the shutdown. If the government insists that any capability which can be steered toward cyberattack help is a controlled technology, then every major lab is exposed, because every major lab has shipped a model that can, under the right prompting, describe a known software flaw. Three things are likely to move in the next 48 to 72 hours: a fuller account of the directive's text, including whether the foreign-national language is permanent or a temporary bridge to a negotiated rule; a public statement from the Department of Commerce or another US agency, which so far has not spoken on the record; and a response from other frontier labs about whether they have received similar orders. OpenAI's deployment-safety page, which Anthropic cites as evidence that the same capability is available elsewhere, is now part of the public record, but OpenAI itself has not been asked to act on it.
The TLDR digest's June 15 roundup also covered two unrelated product items, Z.ai's GLM-5.2 release and OpenRouter's Model Fusion product page, but neither is bound up with the directive and they sit alongside the Anthropic story only as calendar items.
For now, the record is plain: two of the most capable AI models in commercial circulation are offline, the federal order is on the page, and the only major lab that has published a detailed response says the rule that put them there would, applied to the rest of the industry, freeze the frontier.