Anthropic has switched off two of its most powerful AI models, Fable 5 and Mythos 5, after the US government ordered it to block every foreign national from using them on national-security grounds.
The company said in a statement it received the directive on 12 June and complied at once. Because it cannot reliably tell foreign users from American ones in real time, it disabled both models for all customers worldwide rather than risk breaching the order.
Anthropic said the rest of its models, including the wider Claude line, are not affected.
The order covers “any foreign national, whether inside or outside the United States,” Anthropic said, including the company’s own overseas staff. Customers outside the US lost access straight away.
The trigger, according to the company, was a government finding of a way to “jailbreak” Fable 5, which Anthropic described as a narrow flaw tied to asking the model to read a codebase and fix software vulnerabilities.
For readers who do not follow AI as closely as this nerd does: a “jailbreak” is a trick that gets an AI model to ignore its own safety rules.
Chatbots like Fable and Mythos are built with careful guardrails that make them refuse certain requests. A ‘jailbreak’ is a carefully worded instruction that slips past those guardrails and coaxes the model into doing something it was built not to do.
Anthropic says the method here was narrow, and that it turned up only minor, already-known software flaws that other widely available AI models find without any trick at all.
According to reporting, the directive came from the US Commerce Department, which used national-security export-control powers to bar the company from giving foreign nationals access. It is believed to be the first time export controls, a tool built for the likes of advanced chips and military technology, have been turned on a commercially deployed AI model.
Anthropic made clear it disagrees with the decision even as it complies.
“We disagree that the finding of a narrow potential jailbreak should be cause for recalling a commercial model deployed to hundreds of millions of people,” the company said.
Both models are only days old. Anthropic released Fable 5 on 9 June.
Fable 5 is the most capable model the company has yet put in front of the general public, billed as a leap forward in software coding, research and analysis, with safety filters that pass sensitive requests to an older, more tightly controlled Claude model.
Mythos is the same underlying system with some of those guardrails lifted for vetted users such as cybersecurity professionals – including the Australian government, which had just been allowed access to test it.
The thing the US government and its eccentric, orange tsar are most afraid of? That foreign powers would use Mythos and Fable 5 to design their own AI-powered super weapon to exploit holes in code that may have existed for decades but never been found until now.
A.I. is the worst thing ever invented by human