Pentagon wins Anthropic supply-chain case

The ruling: safety limits become a supply-chain issue

The Pentagon has won a key appeals-court fight over whether Anthropic’s Claude can be treated as too constrained for military supply-chain purposes. On September 25, 2026, the U.S. Court of Appeals for the District of Columbia Circuit denied Anthropic’s petitions for review in Anthropic PBC v. United States Department of War, a consolidated case over the government’s decision to exclude Claude from the Department’s supply chain under the Federal Acquisition Supply Chain Security Act of 2018 [1].

The 2-1 majority, written by Judge Gregory Katsas and joined by Judge Neomi Rao, held that the Department had enough support to conclude that continued integration of Claude into its systems, either directly or through contractors, presented a covered national-security supply-chain risk [1]. Judge Karen LeCraft Henderson dissented, arguing that the statute should be read more narrowly and aimed at subversive or deceptive interference, not openly disclosed safety restrictions [1].

At the center of the dispute were Anthropic’s restrictions on Claude’s use for lethal autonomous warfare and domestic surveillance. The court described the Department’s position as a demand for an “all lawful uses” term, while Anthropic maintained that the two remaining limits reflected its safety mission and concern about military and civil-liberties risks [1]. Bloomberg Law reported that Anthropic had sought assurances that its technology would not be used for mass surveillance of Americans or autonomous weapons deployment, while the government wanted to use Claude without company-imposed restrictions [2].

Why the court sided with the Pentagon

The majority’s logic is important because it did not require a finding that Anthropic was hostile, malicious, or secretly sabotaging military systems. Instead, it treated Claude’s built-in behavioral constraints as potentially relevant to the statutory definition of supply-chain risk [1]. The court emphasized that Anthropic can influence how Claude responds to prompts through training and future model versions, and that the record showed instances where Claude refused or failed to process government requests [1].

The court pointed to government concerns that Claude could be “subject to manipulation” in a way that inhibited the Department’s use of the model, including the possibility that a defense system might fail to engage as intended [1]. The majority also cited examples involving refused government queries, including prompts connected to classified-material evaluation and CDC disease-prevention research, as evidence that Anthropic’s restrictions were not merely theoretical [1]. Courthouse and legal summaries published after the ruling likewise framed the decision around Claude’s ability to refuse orders or tasks during sensitive government use [3].

That reasoning turns the usual civilian narrative around AI alignment upside down. In commercial markets, a refusal can be sold as evidence that a model is safer, more responsible, and better governed. In a military procurement setting, the same refusal can be read as a loss of command reliability. The court did not say that Anthropic’s values were illegitimate; it said the Department could decide that those values created uncertainty when Claude was embedded in defense systems [1].

The First Amendment claim failed, but not because safety speech was unprotected

Anthropic also argued that the exclusion punished the company for its public advocacy about AI safety. The majority rejected that theory while still acknowledging that Anthropic’s advocacy about safe and appropriate uses of AI is protected speech [1]. The problem, according to the court, was causation: the Department acted because Anthropic refused a contract term the Department considered essential, not because the company had advocated for AI regulation or safety limits [1].

That distinction is likely to matter beyond this case. A vendor may be free to argue publicly that autonomous weapons or mass surveillance are dangerous, but the government may still decline to buy or integrate that vendor’s technology if the vendor refuses terms the agency says are operationally necessary. In the majority’s formulation, this was a procurement dispute with constitutional overtones, not a speech case that overrode the Pentagon’s national-security judgment [1].

The court was also deferential on process. It found that the Department gave Anthropic notice and an opportunity to respond, and it declined to second-guess the Department’s urgency assessment in an area involving national security and rapidly evolving AI capability [1]. The majority accepted the idea that removing Claude from defense systems could not be as simple as flipping a switch, especially where the model had been layered into other applications [1].

The dissent: this stretches “supply-chain risk”

Judge Henderson’s dissent is the warning label on the ruling. She argued that the Federal Acquisition Supply Chain Security Act’s definition of supply-chain risk is better understood as targeting sabotage, malicious introduction of unwanted function, data extraction, or similarly deceptive and subversive manipulation [1]. In her view, Anthropic’s position was not that kind of threat: the company disclosed its restrictions and enforced them as a matter of policy.

The dissent’s practical concern is that the majority gives the government a powerful bargaining tool over AI vendors. If a supplier’s refusal to remove use restrictions can be treated as a national-security supply-chain risk, then the next vendor may face a stark choice: loosen the model’s policies or risk the same designation. Law Commentary’s post-ruling analysis captured this point, noting that the decision preserves the Pentagon’s separate designation while a California ruling continues to block broader restrictions on Anthropic [4].

This is the procurement-liability problem now facing AI firms. Alignment policies are not just product features or brand commitments; they are contractual positions. When the buyer is the military, a refusal policy may be interpreted not as a safeguard but as an external veto over operational decisions. Bloomberg Law reported that Anthropic said it “respectfully disagrees” with the decision and is considering options including further review [2].

Why this matters for AI vendors

The case’s significance extends beyond one Claude contract. The ruling signals that AI companies seeking high-end government and defense work may be asked to make their models more obedient in the name of operational control. A vendor that advertises low-refusal, mission-assured systems may now have an advantage in Pentagon procurement, while one that insists on broad safety carve-outs may have to defend those carve-outs as compatible with military reliability.

That does not mean all guardrails disappear. The majority itself recognized two competing national-security risks: the Department’s fear that an overly constrained model could fail during an important operation, and Anthropic’s fear that an unconstrained model could hallucinate inappropriate targets for lethal force [1]. The ruling’s key move is deciding who balances those risks. For the D.C. Circuit majority, that choice belongs to the President and the Secretary of War, not the AI vendor [1].

For the AI industry, the message is blunt: alignment can reassure civilian customers and alarm military buyers at the same time. Claude did not lose because the court rejected safety as a social goal. Claude lost because, in this procurement context, its refusals looked like uncertainty in the chain of command. In gaming terms, Claude just failed the Pentagon’s persuasion check; in legal terms, the Pentagon has converted model behavior into supply-chain doctrine.

Sources from the last 72 hours

1. [1]Anthropic PBC v. United States Department of War, No. 26-1049Sep 25, 2026, 4:00 AM UTC
2. [2]Anthropic Faces Court Setback on US Supply Chain Risk Label (2)Sep 25, 2026, 7:15 PM UTC
3. [3]Anthropic loses its Pentagon supply chain risk appeal, though judges accept it had no bad motiveSep 25, 2026, 4:00 PM UTC
4. [4]Anthropic Loses Appeals Court Bid to Overturn Pentagon ‘Supply Chain Risk’ LabelSep 25, 2026, 4:00 PM UTC

AI-generated article based on recent web research, then preserved as a dated editorial snapshot.