Anthropic pulls Mythos-class models globally

Anthropic said the United States government has issued an export control directive to suspend access to its latest Claude Fable 5 and Mythos 5 models by any foreign (non-US) national.

The access ban applies to both inside and outside the US, and applies to non-American Anthropic employees as well, with the company saying it must abruptly disable its flagship large language models (LLMs) for all customers to ensure compliance.

Anthropic issued a statement on the US direction saying that while it did not receive any specific detail of the national security concern, it "understands" the US government has become aware of a way to jailbreak Claude Fable 5.

The company characterised it as "a potential narrow, non-universal jailbreak, which essentially consists of asking the model to read a specific codebase and fix any software flaws."

It disagreed that the potential jailbreak constituted a reason to recall a commercial model that has been deployed to hundreds of millions of people.

Claude Fable 5 was released for public access on June 10, and included safeguards to stop it from being used for cyber security, and biological and chemical research.

Mythos 5 meanwhile was only released to specific Anthropic partners under Project Glasswing, without some of the safeguard layer that Fable 5 comes with; the models are otherwise based on the same underlying technology.

Anthropic said it has not received a disclosure of a non-universal potential jailbreak, or bypass of safeguards, that led to harmful results, for Fable 5.

However, Anthropic said it has reviewed a report that it believes is behind the US government's access ban to the models, and will provide more details on it.

It added that prior to launching Fable 5, the company worked with the US government and Britain's AI Security Institute (AISI) and other organisations to "red team" or test the LLMs' safeguards.

Along with a 30-day data retention period for Fable customers for usage monitoring, Anthropic said it adopted a defence-in-depth strategy with the model that aimed to make jailbreaks difficult and expensive to produce.

Nevertheless, Anthropic said it suspects perfect jailbreak resistance is not possible for any model provider.

The company expects lesser safeguard bypasses that can elicit some cyber information in specific circumstances to be found, and in the future, universal, broad jailbreaks as well.