Amazon Led the Push That Took Anthropic’s Fable 5 Offline, Reports Say — Its Own Investor and Cloud Host

The U.S. government order that pulled Claude Fable 5 and Mythos 5 offline worldwide last week did not start in Washington. According to reporting by Axios and The Information, the push came from Amazon — Anthropic’s largest investor, a board-level backer, and the cloud host that serves its models — whose chief executive personally escalated a security finding to the top of the Trump administration.

Amazon CEO Andy Jassy called Treasury Secretary Scott Bessent and other senior officials on Thursday evening, the reports say, submitting an internal test report in which Amazon researchers claimed to have bypassed Fable 5’s safety guardrails with crafted prompts, coaxing the model into producing information usable for cyberattacks. Amazon was not alone: at least five other companies reached senior administration figures on Thursday night and Friday morning with concerns of their own, turning a single complaint into a chorus.

The escalation moved fast. National Cyber Director Sean Cairncross convened a White House meeting, and the administration first tried to persuade Anthropic to withdraw the models voluntarily. When the company refused, Commerce Secretary Howard Lutnick issued an export-control directive at 5:20 p.m. ET that barred access by any foreign national — giving Anthropic roughly 90 minutes to comply. By 10 p.m., the company had switched both models off for every user on the planet, because it could not police the foreign-national line in real time.

Anthropic continues to dispute that the demonstrated techniques amount to a genuine threat, characterizing the jailbreaks disclosed to it as "minor findings" already reproducible with other publicly available models. It has not been the only skeptic. Cybersecurity researcher Katie Moussouris called the government’s reaction "disproportionate," arguing that what was flagged looked more like "Defense Oriented Prompting" than a true jailbreak — the model doing what a security tool is supposed to do, surfacing flaws in code it is asked to examine.

The detail that lands hardest is the conflict of interest. Amazon sits simultaneously as Anthropic’s biggest financial backer, a board presence, the infrastructure provider that hosts its models, and now the party that helped trigger a regulatory takedown of its flagship release. White House AI adviser David Sacks has said the route back is straightforward in principle — Anthropic remediates the cited vulnerability and the administration wants the matter resolved quickly — but the episode leaves a frontier lab’s most powerful product hostage to a late-night phone call from an investor that competes with it.