Trump administration bars foreign access to Anthropic's Mythos and Fable models

The White House moved last week to restrict foreign access to Anthropic's advanced Mythos and Fable AI models after learning that researchers had bypassed their safety systems in ways that could expose sensitive capabilities to adversaries. The action followed a direct alert from Amazon CEO Andy Jassy to Treasury Secretary Scott Bessent about a jailbreak that compromised Fable's guardrails.

Administration officials described the step as necessary to protect national security. Export controls now bar the models from use by non-U.S. persons, including Anthropic employees who are foreign nationals. In response, the company took both Mythos and Fable offline entirely rather than attempt to operate under the new limits.

Anthropic called the identified flaw a narrow issue and said it was not unique to its systems. The firm also claimed it had received prior government approval to deploy Fable. Officials countered that the company had been warned weeks earlier about risks and had distributed the models anyway. One senior administration source said Anthropic repeatedly chose the path that increased exposure rather than reduced it.

The episode has sharpened existing friction between the company and the Trump team. Officials noted that Anthropic has spent months warning about catastrophic AI scenarios while appearing reluctant to accept restrictions on its own products. One described the company's posture as focused on safety only when it does not affect its own operations. Another pointed to multiple external reports from industry players that raised alarms about Fable before the ban took effect.

Chinese AI developers moved quickly to capitalize on the vacuum. Zhipu AI announced it would release its latest GLM-5.2 model as open-source software with no usage restrictions. Shares of its Hong Kong-listed entity surged more than 30 percent on the news, with Wall Street analysts lifting price targets and highlighting the company's positioning in global markets. Rival MiniMax also saw gains after receiving favorable coverage from major banks.

Inside the administration, the episode is viewed as another example of Silicon Valley firms struggling to align with basic security priorities. Sources described repeated communication failures, with Anthropic officials speaking in technical abstractions that did not address the administration's core concerns about foreign access. One official said the company had been given opportunities to adjust its approach and had not taken them.

The result leaves U.S.-developed frontier models unavailable to international users while foreign competitors face fewer barriers. Administration figures argue the controls were applied only after clear evidence of vulnerabilities emerged. Anthropic maintains the response was disproportionate. The disagreement has left the company's most capable systems sidelined for now, with no immediate path back online under current rules.