On Friday 12 June 2026 the US Commerce Department ordered Anthropic to cut off every foreign national from Claude Fable 5 and Claude Mythos 5. Anthropic could not separate users by nationality, so it switched both models off for everyone. Eighteen days later the order was withdrawn.
Update, 23 October 2026: This article originally reported the shutdown as ongoing, and said it could last weeks. That is out of date. Commerce withdrew the controls on 30 June, and Fable 5 came back on 1 July. We have rewritten the story to cover the whole sequence and removed several claims we could not source. The headline now says what happened.
The directive was real, it was sudden, and it was reversed. The unanswered question is what standard a model now has to meet to stay online.
READ ALSO: Satya Nadella Wants an Emergency Brake on Every AI Model
The order
Anthropic's own statement, published on 12 June, says the government's export control directive arrived at 5:21 pm Eastern. It barred foreign nationals, inside or outside the US and including Anthropic's own staff, from both models. Other models, including Claude Opus 4.8, stayed online. The letter did not name the national security concern. Implicator.ai reports that Commerce Secretary Howard Lutnick signed it and addressed it to Dario Amodei, and that licences would have been required for each export. We could not see the letter itself.
Fable 5 and Mythos 5 had launched on 9 June, three days earlier.
The dispute
Anthropic said the government's evidence was verbal and described a narrow, non-universal jailbreak: asking the model to read a codebase and fix software flaws. It argued other public models, including OpenAI's GPT-5.5, can already do this, and that the standard would "essentially halt all new model deployments" across the industry. It apologised to customers for the disruption.
Who raised the alarm is reported differently. Implicator.ai says Amazon researchers reported prompts that elicited cyberattack-relevant information and that the Wall Street Journal tied talks between Amazon's CEO and the White House to the order. Axios reported a claim by another company that it had jailbroken Mythos. Anthropic has not confirmed either account.
The reversal
On 26 June, Commerce let Mythos 5 go to more than 100 vetted US cybersecurity and critical infrastructure organisations. On 30 June its Bureau of Industry and Security withdrew the controls entirely, so no licence is needed to export either model, according to GovInfoSecurity and Implicator.ai. Lutnick confirmed it on X.
Anthropic agreed to detect and address security risks, coordinate with the government on future releases and report malicious activity. According to Anthropic's redeployment post, it added a classifier it says blocks the flagged technique in more than 99 percent of cases, tested by Commerce's Center for AI Standards and Innovation. Blocked Fable 5 requests are rerouted to Opus 4.8, so some routine coding tasks may get weaker answers. Commerce can reimpose the requirement if Anthropic falls short.
Fable 5 returned on 1 July across Claude.ai, Claude Code, Claude Cowork and the Claude platform. Mythos 5 returned only to vetted US organisations at first, and Anthropic says it is working to expand access to its wider Glasswing partners, in the US and abroad. Its newsroom later listed a release called "Introducing Claude Mythos and Fable 5.1", dated 15 September.
Why it matters, including in Nigeria
Anthropic's own statement says the directive barred all foreign nationals from two live models. For African developers who built on Fable 5 in those three days, the rule applied to them directly: they were foreign nationals under the order. Anything routed through the API stopped working with no warning.
It also showed how fast access can disappear: per Implicator.ai, Anthropic took both models offline within about 90 minutes of the order. Teams in Lagos, Nairobi and Accra that depend on one US model should plan a fallback.
The order landed eleven days after Anthropic's reported confidential IPO filing (reported, not announced) and while it was already fighting the Pentagon in court; see our coverage of its [$965 billion funding round](/news/anthropic-overtakes-openai-965-billion-valuation).
What happens next
A Trump executive order created a voluntary process for early government access to advanced models, and federal agencies face an August deadline for security evaluation standards, per Implicator.ai. Anthropic says it will work with Amazon, Microsoft, Google and other Glasswing partners on a shared way to rate jailbreak severity.
Can Anthropic get the model back online?
- The text of the 12 June letter and the 30 June withdrawal letter, which we could not obtain.
- Who first reported the jailbreak, and whether the Amazon account is accurate.
- When Mythos 5 reaches non-US organisations broadly.
- What jailbreak severity threshold now triggers action.
A ban that lasts 18 days still teaches the same lesson: access to frontier AI can change overnight.
