Why the US gov forced Anthropic’s top models offline

TL;DR: On Friday, the US government ordered Anthropic to bar foreign nationals from accessing its two most powerful models, Mythos 5 and Fable 5, over fears their safeguards could be cracked—prompting the lab to pull both entirely just days after their debut. Anthropic previously warned that these models carry serious cybersecurity risks, yet released Fable anyway after assuring everyone that there were guardrails in place. Now, Washington is signaling that it doesn’t trust the company to vouch for the safety of its own AI models.

What happened: Anthropic claims it had only 90 minutes to comply with Washington’s order after it landed at 5:21pm ET on Friday; it chose to take Mythos and Fable down entirely to ensure no foreign nationals had access.

In a statement published that same day, Anthropic said the government had shown only verbal evidence of a “narrow, non-universal jailbreak” of the models’ safeguards, and said it believed the whole thing was a “misunderstanding.” Over the weekend, the company flew senior staff to DC to try to sort out the issue—though as of now, the models are still unavailable for anyone to use.

How we got here: The initial warning about Fable’s security apparently came from one of Anthropic’s own investors and competitors: Amazon. CEO Andy Jassy told the government on Thursday that Amazon researchers got Fable to cough up data that was supposed to be off limits—which could potentially be useful in cyberattacks.

Yet cybersecurity experts aren’t sold on whether these findings warrant such drastic government intervention. Over 40 of them, led by ex-Facebook security chief Alex Stamos, signed a letter warning the ban could backfire by stripping the best defensive AI from the very teams racing to patch vulnerabilities.

The bigger consequences: The fallout from the Fable takedown has also renewed a global convo about how much other countries can—or should—rely on AI models built in the US; the European Commission called the episode reason to “strengthen its technological sovereignty.” And because the ban covers any foreign national, including those who work in the US, Anthropic’s own non-American researchers are now locked out of working on its top models, too.

Bottom line: Anthropic staffers are meeting Trump administration officials today to make their case for restoring access to these models. The government order comes just about a week before Fable was set to leave Claude subscriptions for pricier usage-based billing on June 23. —WK

Also at Anthropic…

• The AI lab just got sued in a federal lawsuit alleging its Claude Max plans—which can cost up to $200 a month—offer less usage than advertised.