Washington invokes national security and forces Anthropic to suspend its most powerful AI

American artificial intelligence flagship Anthropic was forced on Friday by its own government to cut off access to its most powerful models, three days after their launch, with Washington citing a risk to national security in a decision without known precedent.

Anthropic has announced the suspension of its two new models, its very restricted version Mythos 5 and its limited version for the general public, Fable 5, in order to comply with an injunction issued in the evening.

Washington has ordered, under export control, that access to these models be cut off for "any foreign national, inside or outside the United States," including "foreign employees" of Anthropic, according to the company's statement.

Unable to sort its users, Anthropic, already in conflict with the Trump administration, announced it had to "abruptly disable" both models for all of its customers.

The European Union, which gained access to Mythos in early June after several weeks of negotiations, said this latest development further underscored "Europe's need for technological sovereignty".

"We take note of Anthropic's statement and are assessing the situation," said Thomas Regnier, spokesman for the European Commission, which this month unveiled measures to reduce the dependence of the 27 member states on the United States and Asia for key technologies, including AI.

The announcement prompted reactions on Saturday from leading figures in the French presidential race, who warned against an "AI war" and the need for independence from the United States.

According to the American media outlet Axios, the directive comes from Commerce Secretary Howard Lutnick. He reportedly acted after learning that a company had managed to circumvent the safeguards put in place for these models, which are known for detecting and exploiting cybersecurity vulnerabilities with unprecedented speed and accuracy.

When contacted by AFP, the Ministry of Commerce did not respond.

"We dispute that the discovery of a potential workaround justifies recalling a business model deployed to hundreds of millions of people," writes Anthropic, which wants to see the matter as "a misunderstanding."

"If this standard were applied across the entire sector, we believe it would essentially halt all new deployments of cutting-edge AI models," adds the company, which is at the forefront of global competition against OpenAI, Google and the Chinese company DeepSeek.

Anthropic says it is working to restore access to these latest models "as soon as possible", with the others remaining operational.

"Judicial harassment"?

The company has long campaigned for public regulation of AI: in an essay published Wednesday, its boss Dario Amodei argued for a system of mandatory audits of the most powerful models, inspired by civil aviation, which would give the government the power to block their deployment.

But this must be done "within the framework of a transparent, fair, clear, and fact-based legal procedure," Anthropic points out. Friday's directive "does not respect these principles," the company argues.

"I can't determine whether this is judicial harassment (...) or an overzealous security measure. It's simply grotesque," reacted researcher Dean Ball, White House advisor on AI until summer 2025, on X.

"Security and access are both national interests. Tonight, only one of the two had a say," lamented Vilas Dhar, president of a philanthropic foundation dedicated to AI.

In early June, Donald Trump finally adopted optional government control of the most advanced models, making a cautious shift in his administration which until then had been dominated by opponents of any regulation, accused of hindering competition with China.

Launched on Tuesday, Fable 5 is the first publicly released model belonging to the Mythos class, Anthropic's most advanced line. The company revealed the existence of Mythos in early April but restricted its use to a consortium of companies and institutions, the Glasswing project, due to its cybersecurity attack capabilities.

Fable 5 is restricted in sensitive areas such as cybersecurity and the risks of biological and chemical attacks. Its unrestricted version, Mythos 5, is reserved for some 200 companies, organizations, and government agencies, now chosen "in collaboration with the U.S. government," according to Anthropic.

Security, which Anthropic has made a central selling point, is already putting it at odds with the Trump administration. In early March, the Pentagon terminated its contracts with the company, designating it as a "supply chain risk." Anthropic, whose models were the only ones accredited with national security clearance, has filed a lawsuit: it claims it was penalized for refusing to allow its AI to be used for mass surveillance or autonomous weapons.

Valued at nearly one trillion dollars, Anthropic announced in June that it had filed for an initial public offering, just like its rival OpenAI.

AFP