White House blocks Anthropic’s AI models, raising questions about security and retaliation

Anthropic released a weakened public version of its powerful AI model Claude Mythos on Tuesday, but by Friday both the public version and the original model were cut off for users outside the company. The move did not follow a technical failure or an abuse incident. Instead, Anthropic said the White House issued an unprecedented export restriction order, citing national security.

The company had unveiled Mythos in April, highlighting strong cybersecurity abilities, especially its skill at scanning software code and finding vulnerabilities. Anthropic initially limited access to about 50 trusted organizations, including Google, Microsoft, Amazon, Apple, Nvidia, Palo Alto Networks, JPMorgan Chase, and the company that runs the New York Stock Exchange. In early June, it expanded access to 150 more organizations, including some outside the United States. It also developed Fable, a public version with added safeguards, which launched last Tuesday.

Fable was designed to keep the model from being used in high-risk areas such as biology and chemistry, including weapon planning, and to restrict use in building other large language models. Mythos, by contrast, was made intentionally worse in some advanced-model tasks through prompt and parameter changes, prompting criticism from researchers who said it would be deceptive and harmful to research. Anthropic said it received the government demand on Friday at 5:21 p.m. Eastern time and later said the administration believed Fable 5 could be bypassed through jailbreaking.

Anthropic argued the government did not provide specific national security details and said the jailbreak demonstration identified only a few minor software weaknesses, which other public models could also find. Critics, including security executive Katie Moussouris, called the move an overreaction. The restriction also looked inconsistent with the Trump administration’s easing of some limits on AI chip exports to China. Anthropic is already suing the administration over Pentagon access, after it refused a demand to allow Claude for any lawful purpose and was then removed from federal systems. The dispute remains in court, and the company warned the export ban could signal broader pressure on firms that resist the White House.