Pentagon Says It Dropped Claude. Insiders Say Otherwise

On October 5, a Defense Department official told the BBC that the Pentagon "has ceased the use of Anthropic products." In the same report, former defense officials and contractors said Claude was still running inside Palantir's Maven Smart System the week before, including in operations against Iran. Both things can be true at once, and that's what makes the Anthropic–Pentagon fight worth paying attention to. Most coverage asks whether the Pentagon really stopped using Claude. I think the better question is what happens to an AI lab that won't take its guardrails off.

Why Did the Pentagon Ban Anthropic's Claude?

It started with a contract term. The department, which now calls itself the Department of War, wanted AI vendors to accept "any lawful use" of their models. Anthropic agreed to almost all of it. It drew two lines: Claude would not be used for mass domestic surveillance of Americans, and it would not power fully autonomous weapons that take humans out of the targeting and firing decision.

On February 26, Dario Amodei published Anthropic's statement saying the company "cannot in good conscience" drop those safeguards. In it, he wrote that the department had threatened to remove Anthropic from its systems, to label it a "supply chain risk," and to invoke the Defense Production Act to force the safeguards off. He also pointed out the contradiction: you can't call a company a security threat and also say its product is essential to national security. The next day, Defense Secretary Pete Hegseth announced the designation and gave the department six months to phase Claude out. The formal letter arrived in early March.

A label built for foreign adversaries

The detail that matters most is the label itself. "Supply chain risk" is meant for vendors that could sabotage or subvert US systems, which in practice means companies tied to hostile governments. Anthropic says it had never been applied to an American company before. Using it against a San Francisco lab over a contract dispute turned a procurement tool into a punishment, and it didn't only cover the Pentagon's direct use. Defense contractors were barred from using Claude in Pentagon work as well.

Anthropic's Lawsuit: One Win, One Loss

Anthropic sued on March 9, filing in two courts because the government had issued two separate designations. The results have pointed in opposite directions.

On August 28, US District Judge Rita Lin in San Francisco ruled that one designation was illegal. She found First Amendment retaliation, a due process violation and arbitrary decision-making under the Administrative Procedure Act. She was blunt about it: "the empty invocation of national security is not a blank check to punish and retaliate." She also found the government acted out of a desire to make an example of Anthropic for its "arrogance" in criticizing officials.

Four weeks later, on September 25, the D.C. Circuit went the other way. In a 2–1 decision it upheld the parallel designation, reasoning that Claude's "built-in restrictions" and the unresolved contract dispute could compromise military reliability. Judge Karen LeCraft Henderson dissented. She argued the law doesn't treat "a contractor's honest and upfront enforcement of restrictions" as a supply chain risk. Anthropic says it is "considering all options."

Read the appeals court's reasoning slowly

The majority didn't say Claude was compromised, leaky or foreign-controlled. It pointed to the restrictions themselves. The guardrails were treated as the risk. If that reasoning holds, any lab that says no to a use case can be put in the same category as a vendor suspected of planting backdoors. That's the real precedent here, and it reaches well beyond Anthropic.

Was Claude Still Used in Maven After the Deadline?

By the BBC's sourcing, yes. Hegseth's six-month clock ran out in late August. The October statement came roughly five weeks after that, and even then, multiple people familiar with the matter said Claude was still being used for research, analysis and intelligence work, and in operations against Iran, through Maven, as recently as the previous week.

The reason is unglamorous. Claude was built deeply into Maven, Palantir's system for working through huge volumes of intelligence data such as satellite imagery and drone footage. Lauren Kahn of Georgetown's Center for Security and Emerging Technology told the BBC that these systems "are not just plug and play," and that once integrated, "it can be painful to remove them." I can't confirm whether the October 5 statement is false or just timed to the moment the switch finally flipped. Either way, the Pentagon spent months calling Claude a risk while depending on it in a live war.

What the gap between the statement and the sources tells us

• The ban was never about performance. Nobody kept a dangerous tool running in combat for an extra five weeks. They kept a useful one running because it worked.
• The guardrails didn't stop the mission. Claude did intelligence and analysis work all along. The red lines covered autonomous kill decisions and domestic mass surveillance, not the targeting support that went on through Maven.
• Competitors gained ground. Tools from OpenAI, Google and xAI have spread more widely across the department as Claude was pulled out.
• Switching costs are now a policy weapon. If ripping out one model takes more than six months, labs will think hard before giving the government a reason to try.

The Real Story: What Guardrails Cost an AI Lab

Look at the incentives this creates. Anthropic was the first frontier lab on classified networks and the first to build custom models for national security customers. It wasn't a pacifist holdout. It said yes to nearly everything and no to two things. For that it got a label meant for adversaries, a forced phase-out, a split court record and lost government business to rivals who signed the broader terms.

Every other lab is watching. Safety policies only mean something if a company keeps them when a big customer pushes back. This case puts a price on keeping them, and the price is high. I don't think the lesson most boardrooms take away will be "hold the line." It will be "don't write the line down in the first place."

That would be the worst outcome, and the Pentagon doesn't come out of it better off either. A military that punishes vendors for saying where their tools shouldn't be trusted ends up hearing about those limits less often. Judge Lin and Judge Henderson both saw this. Whether the higher courts agree will decide whether an AI lab's usage policy is a real commitment or just a draft that lasts until the government pushes back.

The D.C. Circuit majority didn't find that Claude was compromised. It treated Claude's built-in restrictions as the risk, and that standard reaches every AI lab with a usage policy.

Key Takeaways

• The Pentagon told the BBC on October 5 that it has stopped using Anthropic products, about five weeks after its own late-August phase-out deadline.
• BBC sources say Claude was still being used through Palantir's Maven, including in Iran operations, the week before that statement.
• Anthropic was labeled a "supply chain risk" in February for refusing to allow mass domestic surveillance or fully autonomous weapons.
• The courts are split: a San Francisco judge called one designation illegal retaliation, while the D.C. Circuit upheld the other 2–1.
• The lasting damage is the precedent. Keeping guardrails can now get a US AI lab treated like a foreign adversary.

Frequently Asked Questions