Saturday Jun 13
GOVERNANCEAnthropic

US Shuts Off Anthropic's Best AI Models

13JUN
ANTHROPICUNPLUGGED

Anthropic's safety pitch just backfired. The US government ordered Claude Fable 5 and Mythos 5 offline worldwide, citing national security. It's the first time Washington has pulled a frontier model.

The directive landed Friday at 5:21pm ET. It hit every user on Earth, not just the foreign nationals it nominally targeted. The company complied within hours, then fought back in public.

Here's the irony. Anthropic spent months calling these its most dangerous releases, too risky for open access. That reputation is exactly what drew the order.

The stated trigger was thin: a narrow jailbreak that reads code to spot bugs. The company notes GPT-5.5 does the same. Days earlier, its CEO asked for this power.

full brief & sources

⚡ Why this matters

  • First time the US government has forced a frontier AI model offline. New precedent.
  • Anthropic's 'uniquely dangerous' safety branding is what drew the scrutiny.
  • Lands two days after Amodei asked Washington for exactly this kind of power.

🔍 What happened

  • June 12, 2026: a Commerce directive forced Claude Fable 5 and Mythos 5 offline worldwide.
  • Framed as export control, but Anthropic says the real concern is a claimed Fable 5 jailbreak.
  • The government gave only 'verbal evidence of a potential narrow, non-universal jailbreak.'
  • Mythos was shared with about 50 vetted orgs - Amazon, Apple, Google, Microsoft - via Project Glasswing.
  • Other Claude models are unaffected; Anthropic is pushing to restore access.

💬 Smart takes

  • Anthropic: recalling a model used by hundreds of millions over a narrow jailbreak would 'halt all new model deployments' industry-wide.
  • Sam Altman, OpenAI CEO (April): Mythos was 'fear-based marketing' - 'we built a bomb, then sell you a bomb shelter for $100 million.'
  • Skeptic: Anthropic says GPT-5.5 has the same capability, so singling out Claude looks arbitrary - or like the safety story working too well.

🧭 Where this goes

  1. Likelyaccess gets restored within weeks after Anthropic's pushback.
  2. Likelyrivals quietly soften their 'our model is dangerous' marketing.
  3. Possiblethis becomes the first test case for Amodei's own model-block framework.
  4. Wild Cardthe standoff dents Anthropic's IPO narrative if it drags.

🥄 The Spoon Take

Be careful what you market. Anthropic spent months calling Mythos too dangerous to release. The government listened and pulled the plug. Two days after Amodei asked Washington for power to block models, Washington used it - on him. Safety-as-moat works until the regulator takes the branding literally.

🤔 Pushback

This may be a short-lived bureaucratic mix-up - access could return in days, making it a headline today and a footnote next week.

·TechCrunch - Government pulls Claude models·Anthropic - Fable & Mythos access·Business Standard - Anthropic halts access