Commerce letter kept Anthropic’s Fable offline for 19 days
Anthropic’s public Fable model, one of the most capable general AI systems on the market, stayed offline for 19 days after a Commerce Department export-control letter forced a takedown, an order that can block foreign access when a product is treated as a national-security risk, Politico Magazine reports in a Sunday reconstruction from more than a dozen participants and contemporaneous documents.
On the evening of April 6, Treasury Secretary Scott Bessent summoned bank CEOs after learning Anthropic would release Mythos the next day to approved public and private users. Late March, Microsoft privately warned after early Mythos access that the model was a significant leap and could automate sophisticated cyberattacks against financial institutions, the energy grid, and other critical infrastructure.
Fable softened Mythos for the general public and landed June 9 after Treasury approval and government pre-deployment testing. Two days later, Amazon researchers stress-testing Fable found a jailbreak, a trick that gets a model past the safety rules the company bolted on.
Amazon CEO Andy Jassy shared the findings with the White House by phone. Amazon is an Anthropic investor; a person familiar said Amazon was responding to an administration request for feedback.
White House aides pressed Anthropic CEO Dario Amodei to take Fable down. Amodei refused to shut it within an hour of first hearing the concerns. That afternoon the lead shifted to Commerce and the Bureau of Industry and Security.
At 5:21 p.m., a Commerce letter directed Anthropic to suspend all access to Fable and Mythos by any foreign national regardless of location, including Anthropic employees abroad. That effectively forced a takedown.
Anthropic said it was complying while disagreeing that a “narrow potential jailbreak” required recalling a commercial model. Access was restored July 1 after 19 days offline.
Commerce announced lifting the Fable export controls on June 30. Negotiations included Defense Department review and Anthropic training an improved safety classifier.
The lab that briefed the government on cyber risk became the hardest fight the White House has had with an AI company.