Pentagon gives Anthropic an ultimatum over AI safeguards

Imagine giving the military the most potent new tool in decades, but telling them they can't touch the trigger without your say-so. That is basically the standoff unfolding between Washington and Anthropic.

Defense Secretary Pete Hegseth has reportedly given Anthropic CEO Dario Amodei until Friday to grant the Pentagon completely unfettered access to its frontier model, Claude. The ultimatum comes with serious teeth: Hegseth threatened to sever the contract, declare the startup a national security supply chain risk, or even invoke the Defense Production Act to force compliance.

Two red lines and a cold war statute

The dispute centers on safety guardrails. Anthropic has stated it is willing to tailor terms for the military, but it refuses to budge on two core red lines: Claude cannot be used for the mass surveillance of Americans, nor can it power fully autonomous weapons that fire without human intervention.

Hegseth made it clear that the Pentagon will not tolerate a private vendor dictating operational decisions or objecting to mission use cases. The tension got personal when Hegseth pointed to an operation in Venezuela targeting Nicolás Maduro, alleging Anthropic had pushed back against Palantir over Claude's deployment. Amodei denied ever raising formal objections and insisted the company's policies have never hindered real-world missions.

If Anthropic holds its ground, the Defense Department's nuclear options could cause immense damage. Invoking the Defense Production Act to strip commercial safeguards from proprietary code would be an unprecedented use of emergency wartime powers. Labeling Anthropic a supply chain risk would be even more damaging commercially, as it would force thousands of military contractors to purge Claude from their daily workflows.

Trapped by Claude's own excellence

The reason the Pentagon hasn't simply walked away comes down to raw performance. Right now, Claude is the only AI model running inside classified military networks.

As one defense official candidly admitted ahead of the meeting, the only reason they are still talking to Anthropic is that they need them right now, and the problem is that they are just that good. In critical military domains like offensive cyber capabilities, Claude reportedly remains comfortably ahead of the competition.

The Pentagon is scrambling for alternatives. Elon Musk's xAI recently inked a deal to bring Grok into classified settings, and officials are accelerating talks with Google and OpenAI to graduate Gemini and ChatGPT to classified operations. But even if Google steps up, it would still face the exact same dilemma: agreeing to let the military use its model for any lawful purpose without reservation.

Why it matters

We think this clash marks the end of Silicon Valley's comfortable illusion that frontier AI can serve state security while keeping clean hands.

For Anthropic, the stakes couldn't be higher. If the Pentagon follows through on declaring it a supply chain risk, the enterprise fallout would be devastating, shutting Claude out of massive swaths of corporate defense tech. Yet caving to Hegseth's ultimatum would permanently gut Anthropic's public identity as the premier safety-first AI lab.

For the rest of the industry, the Pentagon is setting a brutal precedent. If state authorities can wield the Defense Production Act to forcibly override software guardrails, no commercial AI developer truly owns its usage policies. It puts rivals like Google and OpenAI in an impossible position: refuse the Pentagon and risk government retaliation, or comply and accept that your frontier models will power lethal autonomous warfare.