Ceased, Six Months Late: Pentagon Confirms Anthropic Exit as BBC Reveals Claude Ran Iran Operations Until Last Week

On Monday, October 5, 2026, a US Department of Defense official told the BBC, in a statement, that the Pentagon “has ceased the use of Anthropic products.” Six words, buried in a media response — and yet they mark the claimed end of the strangest procurement fight of the AI era: the US military’s eight-month campaign to eject one of America’s own frontier labs from its systems, a campaign that began only because the company refused to remove its safety guardrails.

But the BBC’s reporting contains a far more uncomfortable revelation than the exit itself. Multiple people familiar with the matter — former defense officials and contractors who worked directly on Pentagon AI programs — said that as recently as last week, Anthropic’s Claude was still being used inside the department for research, analysis, and intelligence gathering, including in military operations against Iran. The tools the Pentagon publicly blacklisted in February were, in practice, running live targeting-adjacent workflows deep into September.

What the BBC actually found

The BBC spoke with defense contractors specializing in AI, former senior US defense officials, and others familiar with the Pentagon’s AI use, several of whom spoke anonymously because they were not authorized to talk to the media. Their accounts converge on a simple fact: the phase-out ordered in February did not happen on schedule.

The timeline, as reported:

• February 27, 2026 — Defense Secretary Pete Hegseth designated Anthropic a “national security supply-chain risk,” a legal mechanism normally reserved for companies based in adversary nations. He set a six-month deadline to phase Anthropic out of government entirely. That deadline expired in late August.
• One day later — the US and Israel launched strikes on Iran. From that point forward, the Pentagon’s demand for AI-assisted intelligence analysis surged.
• Throughout the controversy — Pentagon employees continued using Claude, “including its powerful Mythos models,” one source told the BBC. The department’s original August deadline quietly slipped, and it “is not clear why there has been a delay,” as the BBC dryly notes.
• October 5, 2026 — the department finally confirmed the exit, only after the BBC came asking, because multiple interviewees had said Claude was still being widely used.

Anthropic declined to comment on the Pentagon’s statement.

Claude’s actual job inside the machine

The reason removal took months is architectural. Claude was not a standalone chatbot subscription at the Pentagon — it was embedded in Maven Smart System (MSS), the Palantir-operated platform that serves as the Pentagon’s primary system for organizing intelligence, sensor, and visual data. Analysts use Maven to access and organize satellite imagery and drone footage; that data was then fed into Claude and other large language models for tasks such as identifying potential military targets and compiling the one-page briefs passed up the chain of command. Maven is also used by some US allies to share intelligence with Washington.

Sources described Claude as a critical tool for sorting through vast quantities of information. Lauren Kahn, a senior research analyst at Georgetown’s Center for Security and Emerging Technology and a former defense official, put it plainly: because Claude was deeply integrated into Maven, it would have been difficult to remove quickly.

“The fact that it took the Pentagon until now to remove all use of Claude from across the system is indicative that these things are not just plug and play,” Kahn told the BBC. “Once they become integrated it can be painful to remove them.”

That is the quiet lesson of this story for every enterprise betting on AI: frontier models become load-bearing infrastructure faster than anyone plans for. The Pentagon — with unlimited legal authority over its own procurement, a six-month deadline, and every incentive to comply with its own Secretary’s order — still needed extra months to rip out one vendor’s models. A designation meant to be punitive was absorbed as an engineering backlog.

The backstory in one paragraph

Anthropic’s tools had been used across the US government and military since 2024, and the company was the first advanced AI lab to deploy inside agencies doing classified work. Earlier this year, the Pentagon pressed Anthropic to strip Claude’s safety guardrails and grant the military unfettered access. Anthropic refused, citing concerns about mass surveillance and autonomous weapons. Hegseth responded with the supply-chain-risk designation; Anthropic called the move “unprecedented and unlawful” and sued the Trump administration to overturn it. While that fight played out in court, the Pentagon signed replacement contracts with Google, xAI, and OpenAI — and OpenAI’s tools have been more widely adopted in recent months across some military departments, according to two people familiar with defense AI use. One former senior defense official told the BBC there had been internal “concern with moving beyond Anthropic to OpenAI” when the phase-out was first ordered.

The paradox: warming relations, closed doors

The most surreal detail in the BBC report is the timing. The Pentagon’s confirmation of Anthropic’s full exit arrives just as relations between the company and the White House appear to be warming. Trump met Anthropic CEO Dario Amodei at the White House twice last week alongside other technology leaders. After one meeting, Amodei told reporters he still had concerns about AI risks — “How we address those risks is still under discussion… if we work with the president, we can win safely” — and Trump proceeded to praise him as “great” and “fantastic.”

So the picture on October 6 is a company simultaneously suing the administration, being courted by the president, and formally locked out of the Pentagon’s classified ecosystem — while having just moved Claude for Government to general availability for civilian federal and state agencies. The military door closes; the civilian door opens; the litigation continues.

Why it matters

Three takeaways worth tracking:

1. Blacklists are slower than dependencies. A “supply chain risk” designation is designed to sever ties with adversarial suppliers. Applied to a deeply integrated AI vendor, it took over six months to execute — and per the BBC’s sources, the banned tools kept running in active military operations the entire time, including against Iran. Policy instruments built for hardware supply chains do not map cleanly onto software that is woven into analytical workflows.
2. The guardrails dispute is now settled by exit, not resolution. Anthropic never removed its safety limits; the Pentagon never backed down from demanding them. The conflict ended not with agreement but with replacement — the military simply moved its frontier-model business to vendors whose terms it prefers. Whether the replacements operate under equivalent usage restrictions is exactly the question nobody in the department is answering on the record.
3. OpenAI is the structural beneficiary. With Claude gone from Maven and military departments, OpenAI’s defense adoption — already amended once in March after user backlash, to ban domestic surveillance of US persons — becomes the de facto frontier-AI standard inside the Pentagon. That concentration deserves more scrutiny than it is getting.

The BBC notes the details of the department stopping all use of Anthropic’s tools have not been previously reported. The statement that ended the saga was not an announcement; it was a confirmation extracted by a reporter’s question — which is itself a measure of how quietly the Pentagon wanted this chapter closed.