Anthropic's Top AI Scientist Locked Out of His Own Company's Most Powerful AI Because He's Not American

Andrej Karpathy, one of the most respected researchers in artificial intelligence (AI), recently became the face of a massive tech-regulation scare at Anthropic. Rumors claimed he couldn't access his own work. The blamed culprit wasn't seniority, but his passport.

The panic started on Friday when the US Commerce Department issued a sudden export control directive. It banned foreign nationals from accessing Anthropic's two newest models, Claude Fable 5 and Mythos 5. However, Karpathy holds an EB-1 green card, making him a "US person" exempt from the ban.

The order forced Anthropic to disable both models worldwide for every customer at 5.21pm Eastern Time. All other Claude models, including Opus 4.8, stayed online.

One Letter, Hundreds of Millions Cut Off

Commerce Secretary Howard Lutnick delivered the directive to Anthropic chief executive Dario Amodei in a letter that did not detail the specific security concern, the company said in a public statement. A source familiar with the matter told Fortune that Anthropic had 90 minutes to comply.

The company couldn't filter foreign nationals from US users in real time, so it shut both models off for everyone. Fable 5, the flagship consumer model, was launched on 10 June. Mythos 5, an unrestricted version available to vetted organisations under a programme called 'Project Glasswing,' went down with it.

A Safety Researcher Locked Out By Safety Policy

Karpathy co-founded OpenAI, led artificial intelligence at Tesla, and ran the education start-up Eureka Labs before joining Anthropic last month. Though hired to lead a core pretraining research team, the prominent immigrant engineer briefly became the internet's fictional face of the sweeping federal ban.

Co-founder Chris Olah and philosopher Amanda Askell were both born outside the US. While Olah's green card status keeps him legally exempt, Askell and other non-permanent residents remain directly impacted. Anthropic declined to confirm individual staff details.

Amazon Flagged It, Anthropic Disputes It

The trigger was a report from Amazon, Anthropic's largest investor, with roughly $13 billion (£9.7 billion) committed. Amazon chief executive Andy Jassy personally alerted Treasury Secretary Scott Bessent that company researchers had used prompts to extract restricted cybersecurity information from Fable 5, the Wall Street Journal reported.

David Sacks, former White House AI czar and co-chair of the President's Council of Advisors on Science and Technology, posted on X that 'a highly credible trusted partner' had identified a jailbreak and that Amodei 'refused' to fix it before the export order.

Anthropic counters that the bypass produced only a small number of 'previously known, minor vulnerabilities' that other public models, including OpenAI's GPT-5.5, can find without any bypass. Katie Moussouris, chief executive of cybersecurity firm Luta Security, reviewed the Amazon paper and told the Wall Street Journal 'it's not a jailbreak.' She called it standard defence-oriented prompting and said the shutdown 'scored an own goal against us.'

A Precedent That Could Hit Any AI Company

The shutdown lands as Anthropic has filed confidentially for an initial public offering valuing it near $965 billion (£718 billion). On Hyperliquid, pre-IPO futures fell 3% on the news. Defence Secretary Pete Hegseth used the moment to remind the public that the Pentagon 'kicked Anthropic out of our building' three months ago over a separate dispute about military access.

For ordinary users, the message is sharper. Hundreds of millions of paying customers, from coding teams to research labs, lost two flagship models overnight on one national security letter, with no public evidence released. If the standard holds, any frontier model can be pulled the same way.

Anthropic says it's working to restore access and calls the situation 'a misunderstanding.' As of Sunday, neither model was back, leaving the entire global user base completely locked out.