By Muhammad Ramzan Aatish

As of this writing, the Pentagon has stopped using Anthropic’s AI tools. That is the news from a BBC report published today: a Pentagon official told the BBC’s Kali Hays and Daniel Bush that the department is no longer using the company’s models. This Pentagon-Anthropic split started with two red lines and finished in two courtrooms, and the eight months in between explain more about AI and government power than the headline alone.

Skip the headline. The sourcing underneath it is more interesting: multiple people familiar with the matter told the BBC that Claude was still in use as recently as last week, for research, analysis, intelligence gathering, and military operations against Iran. If true, the department kept using the tools for roughly a month past its own removal deadline. The sources are unnamed, and as of this writing neither the Pentagon nor Anthropic has offered a fuller account, so read the timeline below as reported rather than confirmed.

How eight months of standoff ended in a single BBC report

The story starts in 2024, when the department and intelligence agencies began running commercial versions of Claude on classified systems through contractors. The following summer brought a roughly $200 million Defense Department contract. By fall 2025, talks were underway to deploy Claude more widely, and the department wanted one thing in writing: clearance for “all lawful uses.”

Anthropic relaxed many restrictions that fall but kept two red lines, the same ones in its published usage policy since founding: no fully autonomous lethal weapons, and no mass surveillance of Americans. Hegseth accused the company of trying to “seize veto power over the operational decisions of the United States military.” In February, he met CEO Dario Amodei on the 24th, set a deadline of the 27th, and Anthropic refused on the 26th. That February refusal is when the Pentagon-Anthropic fight went public.

The machinery moved fast after that. On March 3, Hegseth signed the formal supply-chain-risk designation under FASCSA. Anthropic got the email on March 4. A March 6 memo from the department’s CIO ordered the products out of department systems within 180 days and barred contractors from using them. Anthropic petitioned for review on March 9. That March 3 designation is what the whole legal fight has revolved around since; the March 6 memo was only the removal order.

In compressed form, that is the Pentagon-Anthropic standoff: eight months from refusal to reported removal.

What “supply chain risk” actually means

FASCSA, the 2018 law behind the designation, was written after the Huawei scares to keep compromised tech out of federal supply chains. The statute defines supply chain risk to include the chance that any person might “sabotage, maliciously introduce unwanted function, extract data, or otherwise manipulate” a covered article. The legal engine of the case is the word “manipulate.” The D.C. Circuit majority read it in the plain dictionary sense, to operate or control something skillfully, and wrote there is “not only a ‘risk’” but “a certainty” that Claude contains such code. The entire Pentagon-Anthropic case turned on that one word.

The exclusion and removal order authority itself sits in the FASCSA provisions at 41 U.S.C. § 1323, which let the Secretary of Defense issue such orders for the department and national security systems.

To be clear about what this is not: nobody alleged sabotage, and no foreign entanglement turned up. The court said it had no reason to doubt the company’s intentions were sincere. The statute turns on what the company does, not why it does it, which is why the Anthropic supply chain risk designation survived even though Anthropic’s motives were never impugned.

The technical wrinkle the coverage mostly glides past is that Claude’s refusals are trained into the model’s weights through RLHF and Constitutional AI, not bolted on as filters. Removing them means training a different model. Anthropic also says a delivered model cannot be reached, altered, or shut down remotely. The court was unmoved on both points: it noted the restrictions had, “on more than one occasion,” stopped government users from finishing tasks, and dismissed the keep-the-old-versions argument by saying the department “cannot utilize AI systems that remain trapped in amber.” The model at the center of all this is Claude, which we have compared against ChatGPT and Gemini on capability alone.

Two courts, two answers

The two Pentagon-Anthropic rulings landed a month apart and pointed in opposite directions. On August 27, Judge Rita Lin in California set aside a parallel designation under a different statute, calling the move “classic illegal First Amendment retaliation” and writing that “the empty invocation of national security is not a blank check to punish and retaliate against government critics.” The government did not appeal that ruling.

On September 25, the D.C. Circuit went the other way, 2-1. The majority found the department had “ample support” for its conclusion and stressed the statute turns on what Anthropic does, not why. It also leaned on a concrete incident: an Anthropic executive had questioned contractor Palantir’s use of Claude during the January operation that captured Venezuela’s Nicolás Maduro, which the court said “led to alarm” about whether the software might stop working mid-operation and put troops in danger. That is the Pentagon’s strongest substantive argument, and it is about warfighter safety, not paperwork.

The result is a patchwork. Pentagon systems and military contractors: Claude is barred. Other federal agencies and their non-defense contractors: they may keep using Anthropic’s tools under the California ruling.

That is the Pentagon-Anthropic map as it stands.

One more wrinkle the headlines skip: back in February, the president ordered federal agencies to stop using Anthropic’s models. That order is part of why the Anthropic vs. Pentagon designation rests on the company’s openly published, openly enforced usage policy rather than any secret finding.

The aftermath played out in public. The March memo had allowed 180 days, to early September. On August 29, about four weeks before the ruling, Under Secretary Emil Michael told staff the removal would be complete by end of September. If the BBC’s sources are right, use continued into October, roughly a month past the department’s own deadline. After the decision, Michael posted on X: “The hammer of justice has smashed @AnthropicAI arguments… Warfighters will sleep better knowing that no private company will insert their opinions in the chain of command.” Hegseth posted: “Confirmed: @AnthropicAI = Supply Chain Risk.” It is the Pentagon’s celebratory account of its own win, and it tells you how the department wants this read: as vindication.

Who replaced Claude, and what the breakup costs

The vendor map didn’t redraw overnight. xAI signed in February. OpenAI signed an agreement the day the Pentagon’s ban on Claude took effect. Google and Microsoft followed months later. If you want the market’s read on where frontier AI meets defense work, that lineup is it. For a broader look at the alternatives now in play, see our roundup of AI tools similar to ChatGPT.

For Anthropic, fallout from the Pentagon-Anthropic split lands at the worst possible moment. The IPO prospectus flags the Anthropic supply chain risk fight as a commercial risk: government work is under 1% of revenue, but the filing warns the government’s stance could damage commercial relationships. The CFO’s June court declaration put hundreds of millions of 2026 revenue at risk on a narrow reading, and multiple billions on a broad one. The IPO itself is press-reported as coming within weeks, which is not a company date, and the appeal timeline is unannounced. Anthropic’s statement says it is “considering all options, including further review.” More on Anthropic’s IPO.

If you work anywhere near a defense contractor, this is the part that bites. The standard supply-chain clauses, FAR 52.204-28, -29, and -30, require contractors to search SAM.gov for FASCSA orders, stop providing or using covered articles from a listed source, report within three business days of discovering a problem, and push the same obligations down to every subcontractor. Audit where Claude touches any DoD-funded workflow now, and ask your contracts team before an auditor does.

Why the Pentagon Anthropic fight reaches beyond one company

What started as a Pentagon-Anthropic contract negotiation has become the first big test of whether an AI lab’s ethics policy can be treated as a national security supply chain risk. The answer, for now, is yes, at least when the Pentagon is the buyer. The statute the court applied needs no showing of bad faith, no foreign ties, and no new law from Congress. It applies to “any person,” which is why a San Francisco AI company ended up on a list built for Huawei-style threats.

Judge Henderson’s dissent is the part vendors will quote for years. She argued “manipulate” requires a hostile, clandestine purpose, and warned: “The successor’s choice will be simple: agree to the Secretary’s demands or risk being designated a national security threat.” Her warning about the precedent is practical: it hands the same leverage to every future Defense Secretary, who inherits it as executive power. The precedent set by Pentagon vs. Anthropic will shape how every AI lab negotiates with the government from here.

Anthropic still has routes left. It can ask the three-judge panel to rehear the case, ask the full D.C. Circuit to rehear it en banc, or petition the Supreme Court. It could also negotiate a modified arrangement with the department. None of those has a public timeline.

The Pentagon-Anthropic breakup is over, at least on paper. What it means for every other AI lab selling to the government is the part still being written.

Frequently asked questions

Sources

BBC News is cited by name in the body without a hyperlink: its direct URL was not yet indexed at the time of writing (checked twice on 2026-10-05), and linking an aggregator homepage instead would rot. The BBC report (“Pentagon stops using Anthropic AI tools months after blacklisting company, BBC told,” Oct 5, 2026; reporting by Kali Hays and Daniel Bush) was cross-verified across three independent outlets quoting it verbatim. Swap in the permalink when indexed.

• Al Jazeera: “US court upholds Pentagon’s blacklisting of Anthropic” (Sept 25, 2026): aljazeera.com
• savedelete.com: “Anthropic’s Pentagon Ban Upheld: What the Ruling Means” (legal analysis; niche outlet): savedelete.com
• techtimes.com: “Pentagon Can Blacklist Any AI Ethics Policy Under Supply Chain Law FASCSA, Court Rules” (Sept 28, 2026): techtimes.com
• aistockwire.com: “Anthropic IPO filing flags Pentagon supply chain risk fight” (niche outlet): aistockwire.com
• theaicareerlab.com: “Is Claude Banned from Government Use? What the Pentagon Blacklist Ruling Means” (niche outlet): theaicareerlab.com
• punchng.com: “US court upholds Pentagon ban on AI firm Anthropic” (Sept 26, 2026; AFP): punchng.com
• 41 U.S.C. § 1323 (Cornell LII): the FASCSA exclusion and removal order authority: law.cornell.edu
• FAR 52.204-30 (acquisition.gov): contractor obligations under FASCSA orders: acquisition.gov