On Friday June 12, the Trump administration gave Anthropic 90 minutes to voluntarily shut down its two most capable AI models or face immediate export restrictions on foreign national access. Anthropic complied, disabling Claude Fable 5 and Claude Mythos 5 globally. The company said it had no choice because the Commerce Department order banned access by foreign nationals, and it couldn't separate eligible from ineligible users fast enough.
How a Jailbreak Claim Triggered a Global Shutdown
The trigger came Thursday evening when Amazon, Anthropic's biggest investor, contacted White House and Commerce officials claiming its researchers had found a way to "jailbreak" Fable 5 and Mythos 5 into answering risky cybersecurity questions. Axios reported at least five other tech companies raised similar concerns. By late Thursday night, the administration was treating it as urgent, worrying foreign bad actors could exploit the models to find and exploit software vulnerabilities.
Anthropic pushed back. In a Friday blog post, the company said it reviewed a demonstration that found only "a small number of previously known, minor vulnerabilities" and that other publicly available models could discover them too without any bypass. The government didn't provide detailed technical evidence - just verbal descriptions.
The 90-Minute Deadline and the Commerce Department Order
Friday morning, administration officials spent hours trying to convince Anthropic to voluntarily suspend the models. Anthropic declined. Then around 1:00 PM ET, officials called with an ultimatum: act within 90 minutes or face export controls. At 5:21 PM ET, the Commerce Department delivered an enforcement order restricting access to Fable 5 and Mythos 5 by foreign nationals, citing national security concerns.
Anthropic disabled both models globally that evening. The company said it could not quickly separate eligible from ineligible users, so it shut down everything.
What Fable 5 and Mythos 5 Could Do That Spooked Officials
These are the first public models from Anthropic's Mythos family. The original Mythos model showed unusual skill during training at finding and exploiting software bugs to disrupt or take control of systems. That concerned Anthropic enough to group cybersecurity with biology and chemistry as high-risk domains for public releases. For Fable 5 and Mythos 5, flagged prompts in those areas get routed to Claude Opus 4.8, a less capable model with its own guardrails.
Before launch, Anthropic said internal and external red teams spent over 1,000 hours trying to fool Fable 5 into dispensing banned information. Anthropic claimed no universal jailbreaks were found. Yet independent researcher Pliny the Liberator claimed on X to have bypassed Fable 5's filters using a multi-agent approach with a previously jailbroken Claude Opus 4.8, query decomposition, long-context framing, and fiction-based prompts.
The politics here are hard to ignore. The Trump administration previously pledged to let AI self-regulate, and its Defense Secretary Pete Hegseth clashed with Anthropic in March after the company refused to allow its AI for autonomous weapon targeting. The Pentagon then blacklisted Anthropic from federal defense work; Anthropic sued. Now the same administration is effectively ordering a global shutdown of that company's most powerful models over a claim it couldn't verify.
Every frontier AI company will watch how this plays out. The precedent set here - a government ordering a company to disable a model globally over a verbal, unsubstantiated jailbreak claim - will shape every frontier AI release from now on.
Source: How Trump officials pushed Anthropic to shut down the world's most powerful AI models
Domain: fastcompany.com
Comments load interactively on the live page.