The U.S. government ordered Anthropic to shut off Claude Mythos 5 and Fable 5 for every user, everywhere, after the model exposed zero-days across every major OS and browser.
Mythos 5 found flaws in every operating system it touched. The Pentagon just made sure nobody touches it at all.
At 5:21 pm ET on a Friday, Anthropic received the directive. Within hours the company disabled Claude Mythos 5 and Claude Fable 5 for all users worldwide.
Access to other models stayed open.
The reach of the order
The directive does not target foreign nationals. It disables the models for every account, every API key, every deployment. Anthropic said on X that it has complied but believes the government got this one wrong.
This marks the first time the U.S. has forced a commercial AI firm to turn off a product globally rather than restrict its export.
How Mythos exposed the vulnerability
Mythos is Anthropic's most capable model. Previewed in early April, it stayed tightly restricted after it identified flaws in every major operating system and web browser tested.
The company opened Project Glasswing to share the model with roughly 50 vetted organizations, including Amazon, Apple, Google, Microsoft, and CrowdStrike, for defensive cybersecurity work.
Fable 5 shipped three days before the order. It added guardrails blocking responses on cybersecurity and bioweapons topics, but the government still demanded the model shut down.
The legal mechanism
The shutdown rests on a supply chain risk determination under 10 U.S.C. section 3252. The Department of War declared that use of Anthropic products in covered systems presents a supply chain risk. The determination took effect immediately and stays in force until the Section 3252 Authorized Official revokes it in writing.
Anthropic had sued the Department of War and secured a preliminary injunction blocking an earlier presidential directive from February 27, 2026, and the March 3, 2026 supply chain designation letter. The government lost that round. It then issued the new order under separate authority that forces the same result without needing to win in court first.
The supply chain label once applied mainly to foreign adversaries such as Huawei. It now covers a U.S. AI company.
What comes next
The government now has a working template: designate a model a supply chain risk, then require its removal from all systems. Within 12 to 18 months, a formal pre-release review board for frontier models with offensive cyber or dual-use capabilities is the logical next step. Companies will face a choice: submit for federal approval before release or lose access to government and allied markets.
Anthropic will lose ground in cybersecurity AI. Rivals with tighter DoD relationships will take the contracts that previously went through Project Glasswing.
What operators should do now
Any organization that built workflows around Mythos 5 or Fable 5 must assume the models can vanish overnight. Diversify suppliers. Assume every frontier model with offensive capabilities will face federal review before deployment.
The model that was built to find every flaw is now the one no one is allowed to run.
