Anthropic vs Trump administration: Who decides the use of advanced AI?

FOR a company that built its brand on caution, Anthropic has had a remarkably public year of confrontation with the US government.

The latest flashpoint, a government order that forced the firm to pull its most advanced artificial intelligence models offline just days after launch, is really an extension of a months-long showdown over the question: Who gets to decide how powerful AI is used, and on whose terms.

It began when the Pentagon in February 2026 pushed Anthropic to allow “all lawful uses” of Claude in defence contracts.

Anthropic baulked, asking for narrow carve-outs: no fully autonomous weapons and no mass surveillance of Americans. The Department of Defense called that an obstruction of legitimate military needs. 

When talks collapsed, Secretary of War Pete Hegseth designated Anthropic a “supply chain risk”, a label usually reserved for foreign adversaries.

The administration ordered federal agencies to stop using its products. Anthropic sued, arguing the designation was retaliation for protected speech and violated due process.

The case is still working through the courts, with the government defending the move as a legitimate contracting dispute rather than punishment for Anthropic’s views.

Controls on frontier AI

Relations between the White House and Anthropic faced another test in June 2026.

The government issued an export-control directive ordering Anthropic to cut off all foreign nationals, including its own employees, from its newly launched Fable 5 and Mythos 5 models, citing an undisclosed national security concern. 

Because Anthropic had no reliable way to screen foreign nationals from its broader user base on a few hours’ notice, it shut the models down for everyone, including paying US customers. 

News articles since have filled in some detail: Officials reportedly believe someone found a way to “jailbreak” the model’s cybersecurity safeguards.

The directive also followed weeks of administration officials privately discussing sanctions amid what The Washington Post described as an erosion of trust in Anthropic, allegedly tied to concerns about the company’s technology reaching a China-linked firm.

Anthropic’s case is that the punishment does not fit the offence. The company says the vulnerability disclosed to it was narrow and minor, the kind every frontier model has and that rival systems are equally susceptible to, and that no actual harmful outcome was demonstrated.

Pulling a commercial product used by hundreds of millions of people over an undocumented and unreviewed claim, Anthropic argues, sets a precedent that could routinely freeze AI deployment industry-wide. 

Anthropic has also noted the irony of being penalised by exactly the kind of opaque government authority and process it has spent years arguing should be more transparent and rule-bound.

The administration’s case, as best it can be reconstructed from officials’ public statements, is that a frontier model with elevated cyber capabilities is a different category of risk than the abstract principle Anthropic invokes.

Waiting for a fuller public accounting before acting on a credible national security tip is itself a risk the government is not obligated to take. 

Trump administration allies have also pointed to the earlier defence dispute as evidence of a company that has been generally unwilling to bend to government priorities, even as Anthropic disputes that the two episodes are connected.

No clear winners

What makes this dispute hard to adjudicate from the outside is that both sides are arguing from positions that sound reasonable in isolation.

A government withholding the specifics of a national-security concern is acting exactly as governments often do; a company objecting to being shut down without seeing the evidence against it is acting exactly as a defendant should. 

The deeper disagreement, visible in both rounds of this fight, is about who bears the burden of proof when frontier AI’s risks are inherently hard to fully verify in real time: the firm building it or the government trying to regulate something it did not build. 

Neither side has fully answered that question yet, and the unusual structure of this conflict – an industry-leading AI lab simultaneously suing and negotiating with the same administration – suggests neither expects to settle it soon.