The Pentagon ripped Claude out. It took months longer than the order said.

On Monday, a Defense Department official told the BBC that the Pentagon "has ceased the use of Anthropic products." The statement did not say when the use ended, and that omission is the story. Defense Secretary Pete Hegseth designated Anthropic a national-security supply-chain risk in February and set a six-month deadline on February 27 to phase the company's tools out. The deadline landed in late August. But multiple people familiar with the matter told the BBC that Claude was still in use as recently as last week, including in military operations involving Iran, for research, analysis, and intelligence gathering. The Pentagon says the transition is done. The BBC's sources describe a transition that ran well past the order.

Why it took so long is the part that matters to anyone buying AI. Claude was not a chatbot the department could simply log out of. It was embedded in the Maven Smart System, the Pentagon's primary data platform, operated by Palantir, which organizes intelligence data including drone footage and satellite imagery. Analysts had woven the model into workflows that processed vast quantities of information and assembled operational briefs. Lauren Kahn, a senior research analyst at Georgetown's Center for Security and Emerging Technology and a former defense official, told the BBC: "Once these systems become integrated it can be painful to remove them." Her kicker: replacing a model wired into a system like Maven is "not just plug and play." Swapping the model meant migrating workflows, not flipping a switch, and it had to happen during active operations.

The cause of the split is a matter of public record, with both sides' positions on it. Earlier this year the Pentagon pressed Anthropic to remove safety guardrails from Claude and give the military unfettered access. Anthropic refused, citing concerns about mass surveillance and fully autonomous weapons. Hegseth then designated the company a supply-chain risk, a label usually reserved for firms tied to hostile foreign governments. Anthropic called the move "unprecedented and unlawful" and sued the administration. An Anthropic spokesman declined to comment on the Pentagon's statement. The courts have split: a federal judge in San Francisco struck down one designation on August 27, while the D.C. Circuit upheld the other 2-1 on September 25, holding that Anthropic's refusal to allow the two uses falls within the supply-chain-risk definition. One panel's sentence will travel: the statute "turns on what Anthropic does, not why Anthropic does it." Meanwhile the Pentagon has signed contracts with Google, xAI, and OpenAI, and OpenAI's tools have reportedly gained wider adoption in some military departments. There is even a thaw at the top: Donald Trump hosted CEO Dario Amodei at the White House last week.

Note on sourcing: the direct BBC article was not retrievable this morning, so every BBC-sourced detail above is framed as the BBC's reporting, corroborated across six independent mirrors that reproduce it consistently. The Pentagon's statement itself, the February designation, and the court outcomes are confirmed across all of them.

Three lessons for procurement owners, none of them about defense.

First, a vendor's guardrail policy is now a first-order procurement risk. Anthropic's refusal to permit two categories of use cost it the Pentagon. Whatever you think of the merits, the mechanism is general: when you build on a vendor's model, you inherit the vendor's policy positions, and those positions can change what you are allowed to do or whether you are allowed to do it at all. Your contract should name this risk. Ask what happens to your deployment if the vendor's usage policy and your requirements diverge, and get the answer in writing before you integrate, not after.

Second, budget the rip-out. The Pentagon had six months and missed the deadline because the model was embedded in the data platform. Your timeline will not be better. If a model is wired into your workflows, your retrieval layer, your review queues, then "we can switch providers" is a sentence you have not earned. The honest version of vendor diversification is measured in engineering quarters, not API keys. Kahn's line is the one to put in front of your team: not plug and play.

Third, watch the pattern across this week's news. In Sydney, the labs asked to be regulated on breach disclosure (see our analysis). In the EU, OpenAI shipped watermarking because a law required it (see our analysis). In Washington, a court ruled that a vendor's safety stance can be a supply-chain risk as a matter of law. Governments are no longer asking the labs how they would like to be governed. They are writing the rules and, in the Pentagon's case, enforcing them by ripping out the product. Procurement built for the era of vendor self-governance needs an update. The new question is not whether your vendor is safe. It is what happens to you when someone with authority decides it is not.