Anthropic, the artificial intelligence company behind Claude Mythos, released a version of its most powerful AI model to the public on Tuesday – despite having warned it was “too powerful” to be shared widely. Within days, the US government stepped in.
The tool, named Claude Fable 5, is a version of the company’s Claude Mythos, which caused serious concern among technology, finance and government leaders when it was first previewed in April. Fable 5 was released with safeguards and user limitations, but Anthropic acknowledged “releasing a model this capable comes with risks”. It added: “Fable’s capabilities exceed those of any model we’ve ever made generally available.”
“Anthropic blocked public access to its 'too powerful' AI model after US government intervened; CEO to meet commerce secretary.”
By Friday, Anthropic had blocked all public access to the model after the US government prohibited the company from allowing any foreign national access to the technology, citing a potential “jailbreak” vulnerability – an opening that could let users force the AI to do something unintended. Anthropic said it had only received “verbal evidence” of the alleged flaw.
Now the company’s chief executive, Dario Amodei, is set to meet US commerce secretary Howard Lutnick in Washington DC on Monday, according to two people familiar with the matter. The meeting is expected to include more documentation of the alleged issue.
Tensions between Anthropic and the US government have been building. The company is already suing the US Department of Defense over a dispute about how its models can be used. A few weeks ago, a meeting with senior White House officials was described as “productive”, but the latest release has reopened the rift.
When Claude Mythos was first made available for preview and testing in April, Anthropic said it was doing so because the tool was so intelligent it could be dangerous, particularly in its ability to exploit or hack computer systems. Canadian finance minister François-Philippe Champagne told the BBC that the attention on Mythos was warranted, in part, because “it’s the unknown, unknown”.
Despite the lawsuit, US government agencies have been testing Mythos. About 150 groups that had early access will now receive Claude Mythos 5, which lacks limitations on cybersecurity or biology depending on an organisation’s uses. Anthropic said those groups have reported finding more than 10,000 critical security flaws in their systems.
Both Fable 5 and Mythos 5 can work “unattended” on human commands for longer periods than any previous Claude models, the company said. Anthropic’s private valuation has neared $1tn (£747bn), and it is expected to become a public company soon.
It remains unclear whether after Monday’s meeting Anthropic will be able to make Fable 5 and Mythos 5 accessible again.