Pentagon drops Anthropic — and what that means for US AI policy

The Pentagon just proved you can blacklist a domestic AI company

The US Department of Defence has stopped using Anthropic’s Claude models across its entire system. The cutoff, confirmed by a Pentagon official in response to a BBC inquiry on Monday, arrives months after Defence Secretary Pete Hegseth branded Anthropic a national security supply chain risk — a designation typically reserved for companies based in adversarial nations. It is a category usually applied to firms with ties to China, Russia, or other countries the US considers hostile. Applying it to an American company that simply refused to deregulate its product marks a sharp escalation in how Washington is treating the AI industry.

But the real story here isn’t that the Pentagon cut Anthropic loose. It’s how messy the decision turned out to be in practice, and what it reveals about the growing friction between AI safety mandates and military operational needs.

Claude was still being used weeks after the ban was announced

Hegseth ordered the phaseout on 27 February, setting a six-month deadline. Yet multiple sources told the BBC that as recently as last week, Claude was still running inside Maven — the Pentagon’s primary data platform, operated by Palantir — where it processed satellite imagery, drone footage, and intelligence summaries for targeting decisions and senior briefing documents.

The tool wasn’t just present. It was critical. Analysts relied on Claude to sort through vast information flows and produce the one-page briefs that travel up the chain of command. That kind of deep integration doesn’t vanish on a timeline. Removing it meant untangling a system built around it.

This is the first time a major AI company has been effectively forced out of the US defence ecosystem through a procurement ban rather than market competition. It also sets an unusual legal precedent: a domestic company branded a national security risk not for ties to a foreign government, but for refusing to remove safety guardrails.

Why Anthropic said no

The origin of the rupture is straightforward. Earlier this year, the Pentagon pressed Anthropic to strip safety controls from Claude and grant the military unfettered access. Anthropic refused. The company cited concerns about mass surveillance and the risk of enabling autonomous weapons systems — arguments that map onto a broader industry debate about whether AI safety constraints should apply to defence applications at all.

Hegseth responded by designating Anthropic a supply chain risk and launching a lawsuit that the administration filed after the designation.

Anthropic has called the move “unprecedented and unlawful” and sued the Trump administration to overturn the designation. The case is still winding through the courts.

The legal question at the heart of the dispute is consequential beyond this single company. If the government can classify a domestic AI firm as a supply chain threat because of its product choices, the same mechanism could theoretically be deployed against any company that declines to meet defence sector demands — regardless of whether that company has any foreign links or national security exposure.

The political wrinkle

Complicating the picture further, President Trump met with Anthropic CEO Dario Amodei twice at the White House last week — the same week the Pentagon issued its cutoff statement. Trump publicly praised Amodei as “great” and “fantastic,” even as Hegseth’s department enforced a procurement ban against the same company.

Amodei told reporters there were still unresolved concerns about AI risks, but that “if we work with the president, we can win safely.” The contrast between the White House’s outreach and the Pentagon’s enforcement action underscores a split that has long existed in US AI policy: the executive branch wants to move fast, while the defence apparatus wants to control the pace.

This divergence is unlikely to resolve itself quickly. The White House has positioned AI dominance as a central pillar of its economic agenda, and Anthropic represents one of the most prominent American AI firms available for partnership. Meanwhile, the Pentagon’s mandate is risk mitigation — and from that perspective, an AI system that won’t operate without constraints is a liability.

Who wins and who loses

OpenAI has been the primary beneficiary so far. Multiple military departments have adopted OpenAI tools in recent months, filling the gap left by Anthropic’s removal. Google and xAI have also signed Pentagon contracts. The commercial reshuffle is already underway.

For Anthropic, the losses are immediate and strategic. The company was the first advanced AI firm to deploy in classified government work. Losing that foothold means losing influence over how AI shapes defence operations — and losing revenue from one of the most lucrative contracts in the sector. It also signals that the administration views compliance as a condition of participation, not negotiation.

For the Pentagon, the loss is operational. Sources described Claude as embedded in intelligence workflows that span multiple agencies and allied partners. Replacing it means retraining analysts, rebuilding integration layers, and accepting a temporary dip in capability — exactly when the US was in the middle of military operations against Iran.

The transition also raises questions about vendor lock-in. OpenAI’s rapid adoption could cement its position as the default AI provider for defence applications, giving it disproportionate influence over how military AI systems are designed, monitored, and deployed — the very dynamic Anthropic was trying to avoid.

What happens next

The legal battle over Anthropic’s supply chain designation will likely determine the long-term outcome. If a court reinstates the designation or upholds it, the precedent opens the door for future use of national security procurement bans against domestic companies that resist safety deregulation demands. That would fundamentally reshape the relationship between AI developers and the defence establishment.

If Anthropic wins, the Pentagon may be forced to reconsider its approach — or find another way to compel compliance short of a procurement ban. Either outcome will matter for every AI company seeking government contracts.

On the operational side, former Pentagon official Lauren Kahn noted that deeply embedded systems don’t unbolt overnight. “Once they become integrated it can be painful to remove them,” she said. The fact that the Pentagon took months to fully exit Anthropic suggests other agencies may face the same delay when asked to replace AI tools they depend on.

Beyond the courtroom, the episode is likely to change how AI companies approach US government contracts. Firms that previously viewed safety guardrails as a differentiating feature may now see them as a liability — or at least something to negotiate around rather than refuse outright. The question is whether that shift comes from fear of blacklisting or from a broader recalibration of what the defence market will tolerate.

The deeper implication is that the US government is now testing whether it can force AI companies to accept fewer safety constraints by threatening their access to defence contracts. Anthropic’s lawsuit is the first real legal challenge to that theory. Whatever court decides it will set the terms for the next generation of military AI procurement — and the answer will echo far beyond the Pentagon’s walls.