Anthropic Stands Up To Pentagon, Refuses To Remove Safeguards

Current frontier models are not reliable enough, says CEO Amodei.

Anthropic has refused to comply with a military edict to remove guardrails on its AI system, Claude.

In a blog post CEO Dario Amodei lists the ways in which Anthropic has already worked with the U.S. military and intelligence communities – Claude is already used “extensively” in these environments – but says his company cannot “in good conscience” agree to the Department of War’s recent request to remove safeguards on its systems.

This follows the Pentagon, and specifically U.S. Secretary of Defense Pete Hegseth, demanding that Anthropic make Claude available for “all lawful purposes.”

If Anthropic does not do so, Hegseth has threatened to mark it as a supply chain risk; a designation normally reserved for firms from foreign adversaries. It has never been applied to a U.S. company.

Anthropic’s issue comes from two red lines it is unwilling to cross: mass surveillance of U.S. citizens, and the creation of fully autonomous weapons.

“We have never raised objections to particular military operations nor attempted to limit use of our technology in an ad hoc manner,” says Amodei. “However, in a narrow set of cases, we believe AI can undermine, rather than defend, democratic values.”

He points out that the two use cases mentioned “have never been included in our contracts with the Department of War.”

Anthropic already supports the use of AI in defense, and even for surveillance. But domestic use is where it draws the line.

“Using these systems for mass domestic surveillance is incompatible with democratic values,” says Amodei. “AI-driven mass surveillance presents serious, novel risks to our fundamental liberties. To the extent that such surveillance is currently legal, this is only because the law has not yet caught up with the rapidly growing capabilities of AI.”

Anthropic doesn’t even have an issue with autonomous weapons, in theory, but today’s systems are “simply not reliable enough” to drive them.

“We will not knowingly provide a product that puts America’s warfighters and civilians at risk. We have offered to work directly with the Department of War on R&D to improve the reliability of these systems, but they have not accepted this offer.”

Hegseth had given Anthropic a deadline of 17:01 local time today to accede to his demands. In his post Amodei accepts that his message may mean a loss of military contracts, and says, “we will work to enable a smooth transition to another provider, avoiding any disruption to ongoing military planning, operations or other critical missions.”

However, it might not be that easy. Claude is already deeply embedded in U.S. systems, including for weapons development and intelligence. It was apparently pivotal in the military operation the U.S. conducted in Venezuela earlier this year.

Computing Says:

This episode could have really tested Anthropic’s claims to be the most safety-first AI company (not that it’s a high bar), especially after it publicly loosened its core safety principles earlier in the week.

While its previous policy said Anthropic would pause training powerful models if their capabilities outstripped its ability to control them, the new policy is both nonbinding and excludes that promise.

The move was initially thought to mean Anthropic was backing down in its fight with the Pentagon. That has clearly not happened.

The ball is now in Hegseth’s court.

This article originally appeared on MES Computing’s sister site Computing.