Anthropic says so shutting down two AI models it launched earlier this week, Claude Fable 5 and Mythos 5, to comply with an export control directive it received Friday afternoon from the U.S. government, citing national security concerns.
This unprecedented incident is the latest flashpoint between Anthropic and the Trump administration. While the company says the order asked it to suspend access for “any foreign nationals, both inside and outside the United States, including foreign national employees of Anthropic,” it has blocked access for all of its customers to comply with regulations.
Earlier this year, Trump’s Department of Defense labeled Anthropic a “supply chain risk” after manufacturer Claude tried to draw red lines over how the U.S. military could exploit its technology. The designation effectively prohibited government agencies and contractors from using Anthropic’s technology. Anthropic responded by filing lawsuits against the Trump administration.
On Tuesday, Anthropic publicly released Claude Fable 5, a version of the company’s Mythos artificial intelligence model with safeguards that prevent it from answering questions about cybersecurity, biology and chemistry. Prior to the release, which Anthropic conducted in partnership with the U.S. government, the Mythos Preview AI model had a constrained launch in April. The goal was to enable companies and organizations to leverage their powerful cybersecurity capabilities to improve their security and allay concerns that the technology could be used by bad actors to develop powerful hacking tools.
In blog entry on Friday, Anthropic says it received a letter from the U.S. government at 5:21 p.m. EST. “The letter did not provide specific details regarding national security issues,” Anthropic wrote.
“We understand that the government believes it has identified a method to bypass, or ‘jail-break,’ Fable 5,” the company added. “We have analyzed a demonstration of the use of this particular technique to identify a small number of previously known, smaller vulnerabilities. All of these vulnerabilities seem relatively simple, and we have found that other publicly available models are also able to detect them without requiring workarounds.”
In a blog post, the company argued that it had implemented mighty safeguards to reduce the likelihood of misuse of Claude Fable 5. Anthropic also argued that the prison break the U.S. government discovered for Claude Fable 5 was minor and would not make the attacker significantly more risky than with another AI model.
“To date, the government has only provided us with verbal evidence of a potential narrow, non-universal jailbreak, which essentially involves asking a model to read a specific codebase and fix any software bugs,” the company said in its blog post. “We understand that information has been shared with the government about one potential prison break.”
Spokespeople for the White House and the U.S. Department of Commerce did not immediately respond to WIRED’s request for comment.
Anthropic CEO Dario Amodei said in a policy essay earlier this week that he and the company support a fair, tidy and see-through government process that would block the release of risky artificial intelligence models. In a Friday blog post, Anthropic argued that “this action is not consistent with these principles.”