On 25 September 2026, a three-judge panel of the U.S. Court of Appeals for the D.C. Circuit denied Anthropic’s petitions to overturn the Department of War’s decision to remove Claude from its supply chain. The 43-page majority opinion, written by Judge Gregory Katsas and joined by Judge Neomi Rao, is not a finding that Claude is dangerous, that Anthropic broke a law, or that the company’s safety work was a sham. It is a procurement-law decision about the meaning of four words in a statute: “supply chain risk.”

That distinction matters because the ruling has been described in some corners as a court endorsing the Pentagon’s view of Anthropic’s safety restrictions, and in others as a judge rubber-stamping a vendetta. Neither description survives contact with the opinion. The court assumed Anthropic’s motives were good. It said so. It then held that the statute does not care about motives. “At least as applied here,” Judge Katsas wrote, “the statutory definition of a ‘supply chain risk’ turns on what Anthropic does, not why Anthropic does it.”

The case is Anthropic PBC v. United States Department of War, No. 26-1049, consolidated with No. 26-1162. It was argued on 19 May 2026 and decided on 25 September 2026. Judge Henderson dissented. The petitions were denied. A petition for rehearing was filed on 13 October 2026; the government opposed it on 18 November 2026. As of 21 November 2026, it is pending.

The short version

Anthropic builds Claude with use restrictions. The Department of War wanted Claude available for “all lawful uses.” Anthropic refused to relax its prohibitions on lethal autonomous warfare and domestic surveillance. The Department concluded that a vendor that can disable its own product in the middle of a military operation is a supply-chain risk, and it removed Claude from the Department’s systems under the Federal Acquisition Supply Chain Security Act of 2018, or FASCSA.

Anthropic challenged the removal as arbitrary, beyond the statute, and unconstitutional. The D.C. Circuit rejected all three challenges. The court found that the Department had ample support for its conclusion that Claude’s restrictions presented a statutorily covered national-security risk, that less intrusive measures were not reasonably available, and that the Department’s notice and opportunity to respond satisfied due process. It also found that the First Amendment did not apply to the Department’s decision because the record showed the exclusion was caused by Anthropic’s refusal to agree to a contract term, not by Anthropic’s advocacy.

The result is a narrow but consequential win for the government. It does not resolve the separate fight over the Department’s use of a different statute, 10 U.S.C. § 3252, or the broader question of how much control an AI vendor may keep over a model once the government buys access. It does give the Pentagon a tested path to treat those restrictions as a procurement risk, and it gives AI vendors a clear warning about what a government contract can demand.

What the court actually held

The majority’s statutory analysis turns on the definition of “supply chain risk” in 41 U.S.C. § 4713(k)(6). The statute defines the term as the risk that “any person may sabotage, maliciously introduce unwanted function, extract data, or otherwise manipulate” a covered information-technology product “so as to surveil, deny, disrupt, or otherwise manipulate the function, use, or operation of” that product or the information on it.

Anthropic argued that “manipulate” in that list should be read narrowly, to require intentionally subversive or deceptive conduct. The Department argued for the ordinary meaning: to arrange, operate, or control. The court sided with the Department. It reasoned that the statute is addressed to the “function, use, or operation” of the product, which focuses on the effect of a manipulation, not the intent behind it. It rejected the idea that the word “maliciously” in the phrase “maliciously introduce unwanted function” should be projected onto “manipulate.” And it declined to import a mens rea requirement into a procurement statute that imposes no criminal or civil liability.

Then came the line that will be quoted in every summary of the case. The court acknowledged that Anthropic “manipulates Claude’s function, use and operation with noble intentions, whether a principled commitment to personal privacy or a genuine concern about AI safety.” It held that those intentions do not change the statutory analysis. “At least as applied here,” the court wrote, “the statutory definition of a ‘supply chain risk’ turns on what Anthropic does, not why Anthropic does it.”

The court also rejected Anthropic’s argument that the Department had to show that Claude was actually malicious or compromised. The statute covers “any person” who may manipulate a product, not only an adversary. The court said the definition is not limited to foreign companies or to sabotage. A vendor’s own ability to disable or constrain its product is enough if it creates the kind of risk the statute describes.

How the court got there

The factual record is unusually specific for a procurement case. The opinion describes a relationship that soured over two years. The Department had used Claude through contractors. It awarded Anthropic a $200 million AI contract in July 2025. It sought to expand the relationship. Anthropic, meanwhile, kept and tightened contractual prohibitions on uses it considered unsafe. The Department says those restrictions stopped Claude from performing tasks government users requested, and that a dispute arose over whether the prohibitions barred the use of Claude in an ongoing overseas military operation.

The breaking point came in February 2026. On 24 February, the Secretary met with Anthropic chief executive Dario Amodei and demanded that Anthropic accept an “all lawful uses” term by 27 February. On 26 February, Amodei published a public refusal. On 27 February, the Secretary announced his intention to exclude Anthropic from the Department’s supply chain. On 3 March, he made the formal written determination required by FASCSA. The notice went to Anthropic on 4 March. On 6 March, the Department’s Chief Information Officer issued the removal order, with a 180-day outer bound for pulling Claude out of all systems.

Anthropic filed a petition for review on 9 March. The Department provided a supplemental notice on 19 March. On 8 April, the D.C. Circuit denied a stay but expedited the case. On 3 June, the Secretary denied reconsideration. Anthropic filed a second petition on 17 June, and the court consolidated the two cases on 24 June.

Date (2026) Event
24 February Secretary demands an “all lawful uses” term by 27 February.
26 February Amodei publishes a public refusal.
27 February Secretary announces intention to exclude Anthropic.
3 March Formal written determination under FASCSA.
4 March Notice emailed to Anthropic.
6 March Removal order issued, with a 180-day outer bound.
9 March Anthropic petitions the D.C. Circuit.
19 March Supplemental notice.
8 April Stay denied; review expedited.
3 June Reconsideration denied.
17 June Second petition filed.
24 June Cases consolidated.

The court’s deference to the Secretary’s national-security judgment is the spine of the opinion. The majority repeatedly said it would not second-guess the Department’s fact-based assessment of the risk, the urgency, or the availability of less intrusive measures. It noted that the Secretary had concluded that removing Claude from the supply chain was necessary to protect national security and that less intrusive measures were not reasonably available. Anthropic argued that the Department could have used conventional procurement authorities instead of FASCSA, and that the FASCSA label stigmatized the company. The court was skeptical. It noted that Anthropic’s valuation had risen sharply since the determination, citing a Wall Street Journal report that the company had received investment offers valuing it at more than $900 billion, and that the stigma would not obviously be less under a different authority.

The court also rejected Anthropic’s argument that the Department’s delay in providing notice was unlawful. FASCSA allows the agency head to temporarily delay notice when an urgent national-security interest requires immediate action. The court found that the urgency was real: the dispute had come to a head, the Department was planning for a military operation, and it needed to shift to other providers quickly. The fact that the Department allowed Claude to remain for up to six months did not undermine the urgency, the court said, because the six-month period was an outer bound, not a grace period.

The constitutional claims

Anthropic raised two constitutional claims. The first was procedural due process. The company argued that it was entitled to notice and a hearing before the exclusion. The court assumed that some process was due, then found that Anthropic received it. The Department notified the company of the exclusion and its rationale, gave it an opportunity to respond, and later provided a supplemental notice and a reconsideration process. The court also found that any earlier-process error was harmless because Anthropic had already mounted a full challenge and could not show prejudice. “For these reasons,” the court wrote, “we hold that Anthropic has suffered no prejudice from the Secretary’s determination that an urgent national-security interest required Claude’s immediate exclusion from the Department’s supply chain.”

The second was a First Amendment retaliation claim. To succeed, Anthropic had to show that it engaged in protected speech, that the government took a materially adverse action, and that the speech caused the action. The court assumed the first two. It found no causation. The record, the court said, showed that the Department acted because Anthropic refused to agree to a contract term the Department deemed essential, not because of Anthropic’s advocacy. The court walked through the timeline: Anthropic had advocated for safety restrictions for years; the Department did not retaliate; it awarded a $200 million contract in July 2025 and continued negotiating after Amodei’s January 2026 article calling for limits on AI weapons. Only after the negotiations broke down did the Department move to exclude the company.

The court acknowledged the Secretary’s “pungent” social-media post, which denounced Anthropic’s “sanctimonious rhetoric,” “virtue-signaling,” and “Silicon Valley ideology.” It said such rhetoric “seldom provides a sound basis for judging the lawfulness of federal executive action.” But it found that the post, “for all its flourishes,” squarely addressed Anthropic’s refusal to provide the “all lawful uses” authorization. The “nub of this dispute was contractual,” the court wrote, “and the First Amendment did not require the Department to continue a contractual relationship that it viewed as creating a national-security risk.”

The dissent

Judge Henderson dissented on the statutory question. She would have read “manipulate” in context to require intentionally subversive or deceptive conduct, not ordinary operation or control. The surrounding verbs, “sabotage, maliciously introduce unwanted function, extract data”, color the residual clause, she argued. A vendor that refuses to accept a contract term is not doing any of those things. Her dissent accused the majority of giving the Secretary a sweeping blacklist power and of reading the statute in a way that turns a contract dispute into a national-security risk.

The majority anticipated the argument and rejected it. It said the dissent’s reading would import a motive requirement that Congress did not write, and that the statute’s focus on the “function, use, or operation” of the product supports the broader reading. The majority also said the dissent’s use of legislative history was misplaced. The disagreement is genuine, and it is the most likely hook for a rehearing or Supreme Court petition.

What the ruling does not say

It does not say Claude is unsafe. The court assumed Anthropic’s safety motives were sincere. It did not find that Claude had actually malfunctioned, leaked data, or caused a military failure. It found that the Department had a rational basis to worry about a vendor’s ability to constrain the product, and that the statute covers that risk.

It does not say the government can ban any AI model it dislikes. The decision is tied to a specific statutory scheme, a specific contractual dispute, and a specific record. The court emphasized that the Secretary’s determination was supported by the record and that the company had notice and an opportunity to respond. A different record, a different statute, or a different kind of restriction could produce a different result.

It does not resolve the § 3252 case. The opinion notes that the Northern District’s determination that the § 3252 designation was arbitrary does not control the § 4713 analysis. The two statutes have different definitions, different procedures, and different jurisdictional rules. The D.C. Circuit has exclusive jurisdiction over § 4713 challenges, and it said that allocation made it inappropriate to constrain its review based on the district court’s judgment.

It does not decide the First Amendment question for AI vendors generally. The court assumed Anthropic’s speech was protected and that the exclusion was adverse. It found no causation on this record. A vendor that can show the government acted because of its advocacy, rather than because of a contract term, would have a different case.

What it means for AI vendors and buyers

For AI vendors, the lesson is that a government contract can reach the model’s behavior. If a vendor reserves the right to restrict how its model is used, the government may treat that reservation as a supply-chain risk. The vendor can still hold the line. It may lose the contract. The court’s decision makes that trade-off explicit.

For government buyers, the decision confirms that FASCSA is a usable tool, not just a paper authority. The Department built a record, made a written determination, gave notice, and offered a reconsideration process. The court upheld the result. Agencies that want to exclude a vendor will need to do the same work.

For everyone else, the ruling is a reminder that “AI safety” is not only a technical or ethical debate. It is also a procurement question. The same restrictions that a vendor markets as a safeguard can be characterized as a reliability risk by a customer that needs the model to perform. The D.C. Circuit did not resolve that tension. It gave the government the benefit of the doubt.

The practical effect may be broader than the holding. Vendors that want to sell to the Pentagon will think harder about how they write use restrictions, how they document them, and how quickly they can change them. Buyers will ask for more operational control. The next contract dispute will be about the scope of that control, not the existence of the safety policy.

What happens next

The petition for rehearing is pending. If the D.C. Circuit denies it, Anthropic can seek Supreme Court review. The dissent gives the company a ready-made argument that the panel misread the statute. The government’s opposition, filed on 18 November 2026, presumably argues that the majority applied the plain text and that the national-security judgment is entitled to deference. There is no guarantee the Supreme Court will take the case. The issue is important, but the record is idiosyncratic.

In the meantime, the removal order stands. The Department has up to 180 days to remove Claude from its systems. The case has already changed the conversation around government AI procurement, and it will continue to shape the terms vendors are asked to accept.

Methodology and caveats

This analysis is based on the opinion PDF published by the D.C. Circuit and the GovInfo copy of the same document. Page citations in the text refer to the slip opinion. We read the majority and dissent in full, checked the quotations against the PDF, and separated the holding from the court’s background description and from the parties’ arguments. We did not speak to the parties, the court, or the Department. We did not review the sealed or non-public portions of the record, if any. The case is still pending on rehearing, so the procedural posture may change. This is analysis, not legal advice.

What we added: We read the 43-page majority opinion and the 8-page dissent, built a timeline from the court’s own recital, pulled the key statutory language, and separated what the court held from what commentators have inferred. The page-cited quotations and the holding-versus-dicta analysis are ours; the underlying documents are linked below.

Last reviewed: 21 November 2026.

Sources: D.C. Circuit opinion, No. 26-1049 (PDF); GovInfo copy of the opinion; GovInfo docket details; CourtListener docket and opinion.

Have a tip? Email [email protected] or use the contact form and choose Editorial. We do not promise anonymity.