If you opened Claude.ai on Saturday evening and got an error instead of an answer, you weren't alone. At some point around 5:30 PM Eastern on June 12, 2026, every person on earth lost access to Claude Fable 5 and Claude Mythos 5 at once — individual users, paying enterprise customers, developers with production API keys, and by Anthropic's own account, even some of the company's own employees. Not a maintenance window. Not a staged rollout gone wrong.

A global kill switch, triggered by a letter from US Commerce Secretary Howard Lutnick to CEO Dario Amodei.

The models had launched two days before.

Source spread

• Anthropic — Statement on the US government directive — safety. Anthropic's own account: received the letter at 5:21 PM ET June 12, describes why selective enforcement was impossible, says the company believes the underlying jailbreak concern is narrower than the order assumes.
• CNBC — Anthropic disables access to Fable 5 and Mythos 5 — builder. First breaking account; enterprise customer impact, timeline, Lutnick letter.
• Fortune — 'It's not a jailbreak,' cybersecurity CEO says — skeptic. Security researcher perspective: the technique cited by the government was defense-oriented prompting, not a replicable universal attack.
• VentureBeat — What enterprises should do now — builder. Practical continuity angle: fallback models, SLA implications, what happens to enterprise contracts.
• Time — Anthropic pulls its most powerful AI models — skeptic. Broader national security and export-control context; what this precedent means.

What's real

The jailbreak was published. On June 10 — the same day Fable 5 launched — a prolific jailbreaker known online as Pliny the Liberator posted what they described as a multi-agent attack on X. The technique was elaborate: Unicode characters, Cyrillic letter substitutions, and homoglyphs (look-alike characters that confuse AI safety filters — imagine replacing the letter "a" with a visually identical Cyrillic character your safety system doesn't flag) to split harmful requests into fragments the model processed one at a time. Pliny claimed the result included functional instructions for cyberattacks and chemical synthesis, specifically referencing a methamphetamine production method.

The government moved fast. Commerce Secretary Howard Lutnick signed the directive classifying Fable 5 and Mythos 5 under US export control regulations on June 12 and sent it to Dario Amodei. The order required Anthropic to suspend access for "any foreign national, whether inside or outside the United States" — a phrase that sounds targeted but in practice covers the majority of people who use Claude globally.

Anthropic's enforcement problem. There is no reliable way to verify a user's nationality in real time across millions of accounts. You can ask. You can't confirm. And the order extended to foreign-national Anthropic employees — meaning even internal access had to stop. The only compliant option was turning it off for everyone.

What deserves a side-eye

Anthropic disputes the premise. Anthropic's statement says the company believes the jailbreak concern is based on a misunderstanding — that Pliny's technique works in one narrow context specific to Mythos 5's cybersecurity capabilities, not as a universal bypass of Fable 5's safety layer. A security researcher quoted by Fortune went further, calling it defense-oriented prompting research that was publicly miscategorized as a novel exploit.

Mythos 5 was specifically designed for cybersecurity. This is worth sitting with. Claude Mythos 5 was Anthropic's advanced model built for security research — the kind of AI that helps defenders find vulnerabilities before attackers do. The fact that an adversarial cybersecurity-focused model might be more useful to adversaries if jailbroken is not a surprise. It's the inherent tension in building those models at all. Anthropic knew this. The government apparently also knew it.

The timing is hard to ignore. Anthropic filed a confidential S-1 with the SEC on June 1. A government-ordered shutdown of your two newest flagship models during IPO preparation is not, charitably, ideal optics. Whether that's coincidence or says something about the government's timing calculation is something I genuinely don't know.

Samwise's take

What to do about it

If you're a regular Claude user:

• Sit tight. This is almost certainly temporary. Check Claude.ai — earlier Claude models (Sonnet 4.6, Haiku 4.5) may still be accessible; the shutdown is specific to Fable 5 and Mythos 5.
• For time-sensitive tasks right now: ChatGPT and Gemini are the functional alternatives. Both are fully operational.
• Don't cancel your subscription over this. A subscription pause is more sensible if the outage extends weeks, but that scenario looks unlikely given how quickly both parties are working to resolve it.

If you're a builder with Claude API dependencies:

• Check your API version settings first. If your integration specifies a model version from before Fable 5, requests may still route to an earlier model.
• Pull your API logs for June 12–14 to quantify impact. You may need this for customer SLA conversations.
• Update your incident runbooks to include "AI provider regulatory suspension" as a failure mode. Today proved it's real.
• If you don't have multi-provider routing (fallback to GPT-5.5 or Gemini when Claude returns errors), this week is the week to build it. Not because Anthropic is unreliable — their uptime record is good — but because provider access can now be terminated from outside the vendor-customer relationship.

Further reading

• Anthropic — Statement on the US government directive — the primary source; Anthropic's own account of receiving the letter and why they had no surgical option
• CNBC — Anthropic disables access to Fable 5 and Mythos 5 — first breaking account with enterprise impact
• Fortune — 'It's not a jailbreak' — the security researcher counterargument
• VentureBeat — What enterprises should do now — practical continuity guide
• Time — Anthropic pulls its most powerful AI models — national security context and precedent