Amazon CEO Andy Jassy personally triggered the U.S. government's decision to shut down Anthropic's two most advanced AI models, a chain of events that saw the company's biggest investor go behind its back and hand Washington the kill switch.
Jassy raised concerns directly with Treasury Secretary Scott Bessent and other senior Trump administration officials this week, according to The Wall Street Journal. Amazon's own researchers had found a way to jailbreak Anthropic's Claude Fable 5 using "a series of prompts," extracting information that could aid cyberattacks. The conversations set in motion export controls that forced Anthropic to disable Fable 5 and Mythos 5 for all users globally.
Commerce Secretary Howard Lutnick sent the directive to Anthropic CEO Dario Amodei on June 12 at 5:21 p.m. Eastern. The order required blocking every foreign national from accessing the models, including Anthropic's own non-US employees.
Unable to filter by citizenship in real time, Anthropic shut both models down for everyone. From launch to death: 72 hours. Anthropic released Fable 5 and Mythos 5 on June 9.
The betrayal angle is what makes this extraordinary. Amazon has invested billions in Anthropic, with plans to go north of $25 billion total. AWS is Anthropic's primary cloud provider. Claude runs on Amazon Bedrock. Amazon holds a board seat. And Jassy still went to Washington and effectively called his own portfolio company's product too dangerous to exist.
David Sacks, Trump's former AI czar who now co-chairs the President's Council of Advisors on Science and Technology, offered his own account. Sacks claimed that "a highly credible trusted partner of both Anthropic and the USG came forward with [information about] a jailbreak." He added: "The Admin asked [Anthropic CEO Dario Amodei] to fix the jailbreak or de-deploy the model. Dario refused."
Anthropic's official response downplayed the finding. The company said the vulnerabilities Amazon identified were "previously known" and "minor," noting that other publicly available models "can discover them as well without requiring a bypass." The company had claimed over 1,000 hours of pre-launch red-teaming found "no universal jailbreak" and had restricted Mythos 5 to vetted partners through a program called Project Glasswing.
None of it mattered. An Amazon spokesperson told the WSJ that while it's "not uncommon for governments to seek our counsel on potential security risks," the company does not "share the details of those discussions." The fallout for the Amazon-Anthropic relationship is immediate.
Jassy did not warn Anthropic or raise the issue privately before going to the Treasury Secretary. The company was reportedly planning an IPO. Killing two flagship models a week before potential IPO paperwork is catastrophic timing.
The broader precedent is just as stark: any major AI investor with Washington access can now effectively trigger a government kill switch on a model. Anthropic spent years positioning as the safety-first AI company, called for more regulation, and warned the public about AI risks.
It turns out the danger was its own business partner.
