Commerce Secretary Howard Lutnick gave Anthropic 90 minutes on Friday to disable Fable 5 and Mythos 5 worldwide, invoking export-control authority that by Friday evening also barred the company from distributing either model to any foreign national, including its own foreign-national employees. It’s the first time the U.S. government has used national-security law to pull a publicly deployed commercial AI model.
The mechanics matter, because the precedent matters more.
According to the Washington Post, Anthropic had given the administration a list of 111 organizations cleared for advanced Mythos access. Officials signed off. The roster then quietly grew by roughly 50 additional entities, one of which was a South Korean telecommunications firm suspected of ties to China. Defense, the CIA, and the NSA started pushing for export controls. The Trump administration had already been weighing sanctions for weeks.
Then came the trigger. Last Thursday, Amazon CEO Andy Jassy told senior officials that Amazon researchers had prompted a Mythos-class model into returning restricted cyberattack information. Fortune, citing Semafor, reports the government believed a China-linked group had already used the same jailbreak. By Friday, Lutnick’s letter was out, with no prior notice.
Anthropic’s response, issued the same day, called it a “potential narrow, non-universal jailbreak” and noted that “perfect jailbreak resistance” isn’t currently possible. Senior technical staff spent the weekend in Washington. By Sunday, a letter calling on the administration to reverse course was circulating, signed first by former Facebook chief security officer Alex Stamos.
Stamos’s own framing is the most useful artifact here. “There was a speeding ticket, and they gave Fable the death penalty,” he told the Post. An unnamed administration official offered the counter-read: “They dug their own grave.” Lutnick, per the Post, had spent the prior weeks “presenting himself as Anthropic’s friend.”
Katie Moussouris, CEO of Luta Security, reviewed the third-party paper describing the jailbreak. The technical dispute over whether the vulnerability warranted the response is real, but it’s also beside the point. What happened on Friday is that an export-control regime built for semiconductors and cryptographic hardware was pointed, for the first time, at model weights running in production.
The structural read is straightforward. Anthropic’s pitch to Washington has always been that safety-conscious labs are the natural partners of the national-security state. That pitch survived the access-list overrun. It didn’t survive Jassy’s phone call, because the call arrived through Amazon, Anthropic’s largest investor and its cloud distributor. The company that hosts you is the company that can end you. CEO Dario Amodei now has to rebuild a relationship in which the administration has demonstrated it’ll use the strongest tool available over what Anthropic itself characterizes as a narrow exploit.
White House AI adviser David Sacks now sits at the center of a regime where the deployment lifecycle of a frontier model runs through Commerce.
Sources
- https://www.washingtonpost.com/technology/2026/06/15/how-anthropic-lost-white-houses-trust-then-its-flagship-product/
- https://fortune.com/2026/06/14/how-a-warning-from-amazon-led-the-white-house-to-shut-down-anthropics-mythos-model/
- https://thehill.com/policy/technology/5926417-anthropic-fable-mythos-ai/
- https://time.com/article/2026/06/13/anthropic-fable-mythos-ban-US-security/
- https://www.nextgov.com/artificial-intelligence/2026/06/anthropic-suspends-top-ai-models-after-us-export-control-order/414173/