Court Lets Pentagon Keep Anthropic on Its AI Blacklist
Matthew Leo · September 26, 2026 · Business & Policy
A US appeals court has allowed the Pentagon to keep Anthropic on a military supply-chain blacklist while the company's legal challenge continues.
The September 25 ruling was not a final decision that every part of the Pentagon's case is lawful. It rejected Anthropic's request for immediate relief from a designation that can prevent military contractors from using the company's technology.
The dispute began with Anthropic's restrictions on using Claude for lethal autonomous weapons and mass domestic surveillance. The Pentagon has argued that a supplier able to withhold service or enforce its own use limits can create uncertainty for military operations.
Why the court sided with the Pentagon
The US Court of Appeals for the District of Columbia Circuit ruled 2–1 against Anthropic. The majority gave substantial weight to the government's national-security judgment and accepted that restrictions controlled by the vendor could be considered a supply-chain risk.
That reasoning is broader than a normal software-security dispute. The Pentagon is not claiming that Claude contains known malware or that Anthropic secretly compromised a military network. The concern is that the company may refuse a use, change access or enforce restrictions after an agency or contractor has built the model into its work.
Anthropic says its safeguards are part of providing a reliable system and that the designation punishes the company for maintaining limits it considers necessary. The company said it disagreed with the ruling and was considering further review.
There are two court fights
A separate federal court in California ruled in August that another Pentagon action against Anthropic was unlawful retaliation and lacked required due process. That decision did not automatically remove the designation involved in the Washington appeal because the two cases concern different legal authorities.
The result is easy to misstate. One court rejected a Pentagon action, while another has now left a different restriction in place. Anthropic has not received a final nationwide ruling that clears every designation, and the September 25 decision does not end the litigation.
The contract question comes first
Military buyers need to know exactly who controls a deployed model. A contract should answer whether the vendor can:
- disable the service or a particular capability;
- change its acceptable-use policy after deployment;
- inspect prompts or outputs for prohibited uses;
- refuse an instruction involving weapons or surveillance;
- require a customer to move to a different model after a safety review.
Those questions do not decide whether lethal autonomous weapons or mass surveillance should be permitted. They establish who has operational control when the government and supplier disagree.
A defence department can reduce dependence by running a model in its own environment, negotiating source or weight access, maintaining a tested substitute and requiring notice before a service change. None of those measures removes the need for human authorization and legal review of military uses.
What Canada should take from it
Canada's Department of National Defence and its contractors will face the same issue when they buy access to commercial models. Procurement documents should state which safeguards are mandatory, which restrictions the supplier may enforce and how service continuity works during a dispute.
A vague promise that a model is “secure” or “responsible” does not settle those questions. The buyer also needs to know whether the model runs in Canada, who can update it and what happens if the provider changes its terms.
Mapletechie's review of Anthropic's outside evaluation arrangement raised a related accountability problem: formal oversight only helps when roles and incentives are clear. Military procurement needs the same clarity before a model becomes difficult to replace.
What happens next
Anthropic can ask the full appeals court to reconsider the panel's decision or continue toward a final ruling on the merits. The company is also dealing with the separate California case.
For government buyers, the immediate issue is simpler. If a critical system depends on a commercial AI provider, the contract needs to say what the provider can refuse, what the government can demand and how either side exits without disrupting an operation.
Sources
- US Court of Appeals for the District of Columbia Circuit opinion, September 25, 2026
- Reuters, September 25, 2026
- Ars Technica
Tags: Anthropic, Claude, Pentagon, Military AI, AI policy