A US federal judge has ruled that the Pentagon acted unlawfully when it labelled AI company Anthropic a supply-chain risk after a dispute over military use of its systems.
US District Judge Rita Lin issued a 59-page order saying the government could not impose sweeping punishment principally because the company criticised the administration’s position. The government is expected to challenge the decision, and a separate, narrower case remains pending in a Washington appeals court.
This is not a judgment that Anthropic’s technology is safe, perfect or entitled to government business. It is a judgment about state power: whether officials had a lawful basis for a designation capable of damaging the company far beyond one contract.
The dispute behind the ruling
The confrontation began after Anthropic chief executive Dario Amodei refused to remove restrictions concerning uses including mass surveillance and autonomous armed drones. President Donald Trump and defence secretary Pete Hegseth accused the company of endangering national security and the Pentagon designated it a supply-chain risk.
Anthropic sued, describing the action as retaliation. AP reported that Judge Lin found the measures were intended to make a public example of the company rather than address an articulable risk that it would sabotage its model.
That is a striking difference. A genuine supply-chain assessment normally asks whether a supplier may fail, be compromised, expose systems or create dependence. Punishing a supplier for refusing a requested capability is a procurement conflict. Calling the conflict “risk” does not automatically supply the evidence.
The designation mattered because federal labels travel. Other agencies, contractors and customers may avoid a company even before courts review the basis. Trump had also directed federal agencies to stop using Anthropic products; Lin previously blocked enforcement while the case proceeded.
Who decides the limits of military AI?
Governments understandably resist allowing a private company to dictate defence policy. Elected institutions must make national-security decisions, and relying on proprietary systems creates democratic and operational concerns.
AI suppliers also retain responsibility for what they build and license. A contract does not dissolve engineering judgment, especially where a system may operate at scale, process surveillance data or influence lethal decisions. The difficult question is not whether government or company gets absolute control. It is how law, contracts, testing and human accountability divide that control.
The judgment does not resolve that larger argument. It says the government must use lawful tools and evidence. The Pentagon can set requirements, choose other suppliers and seek legislation. It cannot, on Lin’s reasoning, attach sweeping penalties mainly because a company objected publicly.
OpenAI reached its own Pentagon agreement after Anthropic was punished, AP reported. That comparison will fuel claims that safety positions are competitive disadvantages. Yet companies’ policies and contracts differ, and the public needs detail before declaring one virtuous and another reckless.
The practical safeguards worth watching are specific: audit logs, restrictions on autonomous targeting, human authorisation, independent red-team testing and clear responsibility when outputs are wrong. “AI for defence” is too broad to evaluate; the system, task and chain of command determine the risk.
The case also arrives as major AI companies consider eventual public listings. Legal certainty around government contracts and safety conditions will matter to investors as much as model benchmarks.
Continue with OutOut’s Technology coverage and Explainers.
The OutOut verdict
The Pentagon tried to turn a disagreement about guardrails into a warning label attached to the entire supplier. The court replied that national security is not a magic phrase that causes evidence and statutory limits to evaporate.
Anthropic has won this round, not the whole war over military AI. The harder task remains: building rules that let democratic governments use powerful tools without outsourcing morality to vendors or treating every safety objection as insubordination. A red button is useful. Knowing who may press it is the actual policy.
