The D.C. Circuit’s 2-1 ruling against Anthropic turns on a residual clause, a narrow refusal, and a question about who gets to weigh competing national-security risks โ not on bad motive, and not on a broad vendor revolt.
This article is not legal advice. The decision was 2–1 with a substantive dissent. The court found no bad motive and held that none is required under the provision at issue, and the August 2026 California decision construed a different, narrower statute — it is not a conflicting ruling on the same question.
Not legal advice. The petitions were denied; the Pentagon prevailed.
On September 25, 2026, the United States Court of Appeals for the D.C. Circuit denied Anthropic’s petitions for review in Nos. 26-1049 consolidated with 26-1162, argued May 19, 2026. Judge Katsas wrote for the majority, joined by Judge Rao. Judge Henderson dissented on substantive grounds. The Department of War excluded Anthropic’s Claude from its supply chain, and the court upheld that exclusion under 10 U.S.C. ยง 4713.
The coverage most people read buried the operative fact. The exclusion followed after Anthropic refused to relax contractual prohibitions on the use of Claude for lethal autonomous warfare or domestic surveillance โ the court’s own words. That refusal was narrow. Anthropic had already written a government-specific addendum to its Usage Policy that contractually allowed Claude to design more effective weapon systems, analyse foreign intelligence, and conduct offensive cyber operations. What it refused to relax were two specific categories. This was not a vendor withdrawing from defense work; it was a vendor that had conceded most of the ground and held two lines.
The immediate trigger was uncertainty, not an outright refusal in the event. As the opinion records, recently, a dispute arose over whether the contractual prohibitions barred the use of Claude in an ongoing overseas military operation, leaving the Department uncertain whether Claude would perform as needed and intended. The Department did not need Claude to refuse a task; it needed to not be wondering whether it would.
The key insight: Anthropic’s entire federal procurement position rested on how broadly or narrowly one residual clause โ “or otherwise manipulate” โ written in 2018 would be read in 2026. The dissent puts it plainly, and the majority does not dispute it: whether Anthropic qualifies under ยง 4713 depends entirely on the scope of that phrase. A residual clause is where a statute stores its future.
The refusal that triggered the exclusion was two categories wide, against a policy that had already been rewritten for this customer.
The Structural Read
Read through the Permission Layer framework, this case is a precise demonstration of how government access functions as a contract term rather than a values statement. The First Amendment claim failed not because Anthropic’s public positions on AI regulation were irrelevant, but because the Department excluded Anthropic based on the company’s refusal to assent to a contract term that the Department deemed essential, not based on the company’s support for greater governmental regulation of AI technology.
That reframes the entire vendor calculus. A usage policy is not a statement of principles that exists alongside a commercial relationship; it is a set of contract terms that constitutes the commercial relationship. Declining to drop a clause is conduct rather than expression, and conduct is assessed differently under the First Amendment than speech. The due-process claim failed separately: the Department promptly notified the company of the exclusion and its supporting rationale, and then gave the company a fair opportunity to contest it.
On bad motive โ two things must be stated explicitly. No bad motive was found, and none was required. The majority had no quarrel with the Northern District’s conclusion that Anthropic has acted with no such bad motive in its dealings with the Department. But the majority held that no such bad motive is required to support a designation under the much broader definition set forth in section 4713. The August California decision โ construing 10 U.S.C. ยง 3252, whose definition the majority called much narrower โ does require bad motive, which is why the two rulings are not in conflict. They construed different statutes with different definitions. Never describe them as opposite conclusions on the same question.
D.C. Circuit โ Majority Opinion, Sept. 25 2026
“The Secretary raises the deeply sobering prospect of overly constrained AI models shutting down unexpectedly and thus causing important military operations to fail. Anthropic raises the deeply sobering prospect of unconstrained AI models hallucinating inappropriate targets for lethal military force. Both possibilities present obvious national-security concerns. But in our Republic, it is the President and the Secretary of War who must determine how best to balance the competing risks.”
Read structurally, the ruling does not say the safety control is wrong. It says the vendor is not the one who weighs it. Both hazards are granted as real; the allocation of the judgment is the holding. The dissent’s argument โ substantive, running on the canons of statutory construction around the residual clause โ is that the majority’s reading of “or otherwise manipulate” stretches beyond what the enumerated terms around it can support. Sabotage, maliciously introduce unwanted function, extract data: all describe an adversary acting against you. The catch-all had to carry a case nobody drafting it in 2018 was imagining, and the majority held that it did. The legal objection to that conclusion belongs to the dissent, not to this analysis.
Permission Layer โ Business Engineer Framework
The Residual Clause as Strategic Chokepoint
In the Permission Layer model, government access is not a sales motion โ it is a contracting event. The operative question for every AI vendor with a federal relationship is not what it says publicly but what it will sign. A usage policy’s residual clauses are the terms most likely to be tested at the worst moment, because residual clauses are precisely where drafters place the cases they could not yet name. The Department’s own position in the record framed this without ambiguity: it called Anthropic’s behaviour a textbook case of how not to do business, demanded full, unrestricted access to Anthropic’s models for every lawful purpose in defense of the Republic, and characterised the refusal as an unacceptable veto power over the operational decisions of the United States military.
Three Implications
IMPLICATION 1 โ CONTRACT TERMS ARE THE PRODUCT
For any AI vendor in a federal procurement relationship, the usage policy and its addenda are not marketing collateral or ethical positioning โ they are the contract. Which terms a vendor accepts and which it declines defines the scope of its federal addressable market more directly than its technical capabilities do. The question of what a vendor permits gets asked at signing and answered under operational pressure, and uncertainty about the answer was itself disqualifying here. The Department did not need Claude to fail; it needed to not be wondering.
Anthropic’s entire procurement position rested on the scope of one phrase โ “or otherwise manipulate” โ written in 2018. The enumerated terms around it described adversarial actors; the residual clause had to carry a case its drafters were not imagining. Every AI vendor with government contracts should audit its usage policies not for their enumerated terms but for their catch-alls, because those are the terms that will be read expansively when no enumerated term fits. A residual clause is where a statute, and a contract, stores its future.
IMPLICATION 3 โ THE WEIGHT OF COMPETING RISKS IS NOT THE VENDOR’S TO HOLD
Nothing above is legal advice.The decision was 2–1 and Judge Henderson’s dissent is a substantive argument about the residual clause, not a gesture. The court found that Anthropic acted with no bad motive, and held that no bad motive is required under 41 U.S.C. § 4713’s broader definition. The August 2026 Northern District of California decision is not a conflicting ruling on the same question: it construed 10 U.S.C. § 3252, whose definition the majority called much narrower and which does require bad motive, and the court expressly had no quarrel with it. Quotations above are taken from the opinion (51 pages; argued 19 May 2026, decided 25 September 2026; Nos. 26-1049 consolidated with 26-1162). Rehearing en banc and Supreme Court review are both discretionary, and nothing above forecasts whether either happens or how it would come out. Contract values and dollar figures, the identity of the overseas military operation, the number of government users affected, and any stated intention to appeal are not established and do not appear; nothing above asserts anything about a withheld mandate. Nothing above predicts the conduct of Anthropic, the Department, any other vendor or any other agency.
Gennaro Cuofano is a CRO and tech executive who has worked in AI since late 2015, starting with bringing NLP, natural-language generation, voice and chatbot products to the marketing industry. His background is in law and finance: he holds a Master's degree in Law and an International MBA with an emphasis on corporate finance (LUISS Business School and the University of San Diego, 2012), and worked as a financial analyst at a real-estate investment firm in San Diego and as an assistant controller. His work focuses on business model strategy, business engineering and, more broadly, structural analysis: how companies make money, read from their own filings. He created FourWeekMBA and leads research there and at The Business Engineer, his newsletter on AI and business strategy, with over 95,000 subscribers and more than 1,000 published analyses. His writing has also appeared on Entrepreneur (for example, How to Digitize a Family Business in Europe) and HackerNoon. Find him on LinkedIn and Substack. See how we source.
Scroll to Top
Discover more from FourWeekMBA
Subscribe now to keep reading and get access to the full archive.
