An AI system that posed such a risk that the US banned it is finally returning.
Anthropic, which makes the chatbot Claude, has in recent weeks detailed its work on its most advanced Fable and Mythos models. But, soon after they were launched, officials said that they posed so much of a danger to national security that they should be restricted.
The government was concerned that the systems could be used to find vulnerabilities in key infrastructure and could therefore be used by military intelligence in countries such as China or Russia to launch attacks.
That led to the US government exposing an export ban that stopped foreign nationals from using the tool. But Anthropic said that it would be impossible to comply with such an order without banning it outright, because it would not be possible to check the nationality of users.
Now, Anthropic has said that the ban has been lifted and that it has been working with the US Commerce Department to keep its latest models available. That includes changes to the system to allow its controls to be harder to bypass.
Last week, the US government allowed the company to release Mythos 5 but only to some "trusted" US organisations. The model — designed to detect cybersecurity vulnerabilities — had earlier been made available to a broader group of companies as part of its Glasswing project.
It is working with the US government to expand access to Mythos 5 to the broader set of domestic and international partners in the Glasswing program, it said in a blog post. Fable 5, which is intended for the general public and has stronger safeguards, will be available from Wednesday.
Anthropic, which has had a rocky relationship with Donald Trump's administration this year, also said it was deepening its collaboration with the U.S. government, giving designated government partners expanded early access to both its models.
It is also working with Amazon, Microsoft, Google and other Glasswing partners to develop shared common standards to assess and fix potential AI jailbreaks (techniques that bypass safeguards), including a system to rank the severity of such jailbreaks.
A letter to Anthropic from commerce secretary Howard Lutnick said Anthropic had agreed to work diligently with the US government on protocols for Mythos, Fable and future models and to inform the US government of any malicious activity.
Mr Lutnick, however, said that the department "reserves the right to reevaluate the decisions made in this letter and the necessity of reimposing a license requirement, should circumstances change or should Anthropic fail to adhere to its commitments".
Anthropic said the June 12 export-control order followed Amazon researchers reporting a way to bypass Fable 5's safeguards, allowing the model to identify software vulnerabilities and, in one case, generate code demonstrating how one vulnerability could be exploited.
Anthropic said it has now implemented a new safeguard that blocks the behaviour described in the report.
Any blocked request will instead be sent to its Opus 4.8 model, and Anthropic acknowledged that although this would be frustrating for users, the tradeoff was made in the interest of making the model's other capabilities widely available.
Anthropic warned that it was "probably impossible" to make any AI model fully robust to jailbreaks and noted the potential for the development of a universal jailbreak that would be able to unblock "an entire class of harmful behaviors."
"There will be many minor jailbreaks, some narrow harmful ones, and although no universal jailbreaks for Fable 5 have been discovered at the time of writing, expert safety researchers continue to red-team it," the company said.
Isaac Harris, executive director of the Frontier Security Institute, a nonprofit focused on AI and national security, said there now appears to be a process for standards of the U.S. models.
But he added: "There's still a question mark as to how equivalently dangerous capabilities coming from China with less guardrails will be handled by the administration in the U.S. market."
Increased scrutiny of AI models this month began with U.S. President Donald Trump's signing of an executive order establishing a voluntary framework for AI developers to offer "covered frontier models" to the U.S. government for up to 30 days before releasing them to trusted partners.
Like Anthropic, rival OpenAI has also faced restrictions. It said on Friday that it had delayed a full public launch of GPT-5.6 at the U.S. government's request, limiting its access to a small group of vetted partners.
Anthropic's relationship with the U.S. government has been particularly tense. The Pentagon earlier this year designated the company a "supply-chain risk", preventing contractors from using Anthropic's AI when working for the U.S. military, after the company refused to allow its models to be used for mass domestic surveillance or fully autonomous weapons systems.
Both OpenAI and Anthropic have confidentially filed for U.S. initial public offerings.
Additional reporting by agencies
Read MoreOpenAI to delay IPO after Sam Altman spooked by SpaceX tumble
