Anthropic took its Claude Fable 5 and Mythos 5 AI models offline on June 15, 2026, after the Trump administration imposed export controls over concerns the systems could be jailbroken and concerns about compliance with a recent cyber executive order. An administration official told Axios that the company had been given opportunities yet “screwed us” by failing to address risks adequately, while Amazon CEO Andy Jassy contacted Treasury Secretary Scott Bessent to highlight vulnerabilities. The firm maintained it had received explicit government approval for deployment of one model and had worked to revoke improper access.
The models became available June 9, with Claude Fable 5 serving as the public-facing version of the Mythos architecture that included extra safeguards blocking high-risk tasks in areas such as biology and cybersecurity. Claude Mythos 5, by contrast, offered fuller capabilities but was limited to vetted partners via Project Glasswing, according to Anthropic’s own system card and announcements. Officials learned that access had extended to an entity connected to the Chinese Communist Party, prompting threats of controls about two weeks earlier.
A source close to Anthropic told Axios the company had collaborated with the government on expanding controlled access and addressed issues before formal warnings escalated. The firm also issued a blog post that dismissed elements of the Amazon report on potential jailbreaks. Nevertheless, the administration moved forward with export controls imposed Friday night, which required a worldwide suspension because Anthropic could not segment users by nationality in real time.
This latest dispute continues a series of frictions between the company and U.S. government entities. Anthropic walked away from a $200 million Pentagon contract in February 2026 rather than ease its safety guardrails for military uses, CNN reported at the time. The company has long emphasized responsible development through published core views on AI safety and a Responsible Scaling Policy requiring detailed risk reports for frontier systems.
White House officials had sought to repair ties after the earlier Pentagon disagreement, yet differing communication styles and personality clashes undermined those efforts, the Axios report detailed. Anthropic involved cybersecurity experts the administration viewed as politically misaligned, including individuals celebrated by former officials dismissed by President Trump. The Commerce Department has arranged a meeting for Monday with Anthropic staff members Logan Graham, Dave Orr and Nicholas Carlini to review compliance with the cyber executive order.
The controls mark one of the most significant U.S. government steps so far to limit distribution of the most advanced AI models on national security grounds, an Associated Press account noted. Anthropic’s system card outlined hundreds of pages of safety testing for the new releases. Industry participants continue to grapple with the pace of technological gains against the slower development of regulatory frameworks.
