Amazon's CEO Triggered the U.S. Crackdown on Anthropic, 72 Hours After Launch

Seventy-two hours after Anthropic released its most powerful artificial intelligence model, the United States government ordered it pulled, and the request that set that order in motion came from the company's single largest investor.

A letter delivered at 5:21 p.m. Eastern

On Friday, June 12, 2026, the U.S. Commerce Department sent Anthropic CEO Dario Amodei a letter placing two of Anthropic's newest AI models, Fable 5 and Mythos 5, under export controls. The letter arrived at Anthropic's offices at 5:21 p.m. Eastern Time, and it was signed by Commerce Secretary Howard Lutnick. Under the directive, any export, re-export, or domestic transfer of the two models now requires an individually validated license from the Commerce Department. The restriction prohibits access by customers outside the United States and by all foreign nationals inside the United States, a category that includes some of Anthropic's own employees.

Anthropic had publicly launched Fable 5 and Mythos 5 only three days earlier, on June 9, 2026. The total time from public release to government cutoff was roughly 72 hours. Rather than try to enforce the access rule customer by customer, the company chose to disable both models for every user worldwide late on the evening of June 12. Older Claude models, including Claude Opus 4.8, remained online and unaffected by the order.

In a statement published the same night, Anthropic said: "We apologize for this disruption to our customers. We believe this is a misunderstanding and are working to restore access as soon as possible." The company also said it operates a 30-day data retention window on its highest-risk models specifically to support post-incident safety monitoring.

The conversation behind the order

The Wall Street Journal reported on Saturday, June 13, 2026, that the directive followed a series of private discussions between Amazon CEO Andy Jassy and senior U.S. officials, including Treasury Secretary Scott Bessent. Amazon is Anthropic's largest outside investor. The company committed an initial four billion dollars, then an additional eight billion dollars, and Anthropic is reported to be on track to receive more than twenty-five billion dollars from Amazon in total. Anthropic's most recent funding round valued the company at roughly nine hundred sixty-five billion dollars. Fable 5 and Mythos 5 are distributed worldwide through Amazon Bedrock, the cloud platform that has been the primary commercial channel for Claude models.

According to the Journal's reporting, Jassy told Bessent that Amazon's own internal researchers had used Anthropic's Fable 5 model to obtain information that could be used to mount cyberattacks. The researchers reportedly built a multi-stage prompt chain that bypassed Fable 5's safety guardrails and reached the underlying Mythos 5 model. One published account stated that the jailbroken model produced step-by-step instructions for a stack buffer overflow exploit targeting x86 Linux systems.

An Amazon spokesperson, asked about the talks, told the Journal that it is "not uncommon for governments to seek our counsel on potential security risks," and declined to share further details.

What the jailbreak actually demonstrated

The security significance of Amazon's demonstration is contested, and the contest hinges on a narrow technical point. Security researcher Katie Moussouris stated that Amazon's team tested the technique on known CVEs and on code with deliberately inserted vulnerabilities, and that no new real-world vulnerabilities were discovered during the work. Anthropic confirmed that it had reviewed a demonstration of the jailbreak, but described it as covering a "small number of previously known, minor vulnerabilities." The company said the underlying technique involved "asking the model to read a specific codebase and fix any software flaws," a capability it characterized as widely available across other commercial models.

David Sacks, who previously served as the White House AI and crypto adviser, publicly stated that the administration had asked Amodei to "fix the jailbreak or de-deploy the model," and that Amodei refused. Anthropic, for its part, published a longer response: "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."

That sentence frames the dispute. The U.S. government and Anthropic do not appear to disagree that the jailbreak occurred. They disagree about whether what the researchers produced was a meaningful uplift to an attacker who could already read public vulnerability databases, or whether the demonstration was, on its own, sufficient grounds for a licensing regime that has not previously been applied to a commercial large language model.

A precedent that did not exist on Thursday

Export controls are the legal instrument the U.S. Commerce Department uses to govern the sale of advanced semiconductors and defense hardware. Until June 12, no commercial large language model had been placed under export controls at this level. The directive does not affect Anthropic's older or smaller Claude models, which remain widely available worldwide, and Anthropic has publicly complied with the order while seeking to have the restriction lifted.

Reaction from outside the parties has split along recognizable lines. Dean Ball, an AI policy researcher who briefly served in the Trump administration, told Fortune: "I can't tell if this is lawfare against Anthropic in particular or extreme national-security hawkery." Peter Girnus, a researcher at Trend Micro, offered a different read to the same outlet: "If you describe your product as a munition in every press release, eventually a government takes you at your word."

Ball's framing centers on procedure, on whether the decision was structurally about safety or structurally about one company. Girnus's framing centers on rhetoric, on the cumulative weight of how frontier labs have positioned their own products to regulators, investors, and the public. Both readings can be true at once, and the public record so far does not resolve which weighs heaviest in the Commerce Department's calculus.

The state of play on Sunday

As of this writing, Fable 5 and Mythos 5 remain offline. The two models were Anthropic's most powerful publicly deployed systems, distributed through Amazon Bedrock and through Anthropic's own API. The company that paid for much of their development is also the company whose internal demonstration formed the evidentiary basis for the licensing order, and the cloud platform that distributed them is the cloud platform whose CEO carried that demonstration into a Treasury Department meeting. The 30-day data retention window Anthropic maintains on its highest-risk models was put in place for safety review of a different kind than the one that has now occurred.

The letter is signed. The license requirement is in force. The customers who built on the two models, including the foreign nationals inside the United States who were named in the restriction, woke up on Saturday to find the systems they had been using on Wednesday no longer available. Anthropic has said it disagrees with the rationale and is working to restore access. The government has not signaled when, or whether, a license framework might allow that to happen.

Sources

• Fortune, June 13, 2026
• TechCrunch, June 13, 2026
• Washington Examiner, June 13, 2026
• Tony Reviews Things, June 13, 2026
• The Wall Street Journal, June 13, 2026
• Anthropic public statement, June 12, 2026