US blocks Anthropic AI exports over foreign military intelligence concerns

WASHINGTON: US Commerce Secretary Howard Lutnick ordered Anthropic to halt exports of its latest Mythos and Fable artificial intelligence models after officials concluded they could be used by military intelligence actors in China, Russia or other countries viewed as security concerns.

According to a copy of Lutnick’s letter to Anthropic chief executive Dario Amodei seen by Reuters, the order was sent on Friday and directed the company to suspend exports of the models worldwide and to all foreign nationals regardless of location. Reuters reported that senior Anthropic technical staff met Commerce Department officials in Washington on Monday to try to work out a solution.

A Trump administration official told Reuters the talks were aimed at addressing government concerns that the models could be used in ways harmful to the United States, while Anthropic is seeking to restore access to its most advanced systems after taking them offline for all users on Friday. A source familiar with the matter said Lutnick has remained directly involved and has been holding regular calls with Anthropic officials as discussions continue. Amodei and Lutnick are both expected to attend the G7 meetings in Evian-les-Bains, France, where they may also speak as negotiations proceed.

The same source said National Cyber Director Sean Cairncross took part in Monday’s working-level meeting at the Commerce Department. A person close to Anthropic told Reuters the company’s technical staff have been in virtual meetings with officials almost daily since the administration first contacted the company on Friday.

Government cites safeguard bypass concerns

After receiving Lutnick’s letter, Anthropic said it would disable access to the models globally. In a blog post published on Friday, the company said the government had informed it that officials believed there was a way to bypass, or jailbreak, a safeguard meant to stop Fable 5 from being used to identify software vulnerabilities.

Anthropic said the bypass exposed only "minor" security flaws and argued that other publicly available models were also capable of finding them.

Reuters reported that relations between the Trump administration and Anthropic had already deteriorated earlier this year after the company declined to let the US military use its AI systems for domestic surveillance and fully autonomous weapons systems. The government then placed the company on a national security blacklist.

The San Francisco-based startup, which has confidentially filed for an initial public offering in the United States, had earlier raised concerns about the hacking potential of its Mythos model and had not broadly released it. On June 9, Anthropic launched a public version called Fable 5, which it said included cybersecurity protections. A person close to the company told Reuters that Anthropic had worked with the government to test Fable 5 before release and had received approval to deploy it.

Rare export control move raises legal questions

Lutnick’s letter said the Commerce Department was acting under powers granted by the 2018 Export Control Reform Act to impose controls on emerging technologies considered vital to US national security. An export control expert told Reuters this was the first time the department had used those authorities in this way.

The letter said a licence would now be required for export of the models or for transfer to a foreign national inside the United States. It also warned that any failure to comply would trigger "prompt criminal and civil penalties."

However, export control specialists told Reuters that AI models are usually not exported in the conventional sense. Instead, they are made available remotely, and current export regulations do not typically govern that form of access. That has raised questions over whether the Commerce Department has the legal authority to impose such restrictions. Reuters said the department did not answer questions about the legal basis for the action, and neither Anthropic nor the department responded to requests for comment on Monday’s meeting.

Reuters also reported that more than 80 cybersecurity executives and experts signed an open letter on Sunday to Lutnick and Cairncross backing Anthropic’s position. The signatories, including cybersecurity leaders from Nvidia and Adobe, urged the Trump administration to remove the restrictions on the company.