Judges Say the US Can Blacklist Anthropic Over Withheld AI Features

Judges decided the US had authority to blacklist Anthropic technology for the company's withholding of certain AI features, and that the authority holds even if Anthropic had no malicious intent. In other words, a vendor's judgment about which capabilities to hand over does not shelter it from being cut off from government work, provided the government's action stays inside its statutory and constitutional limits — which, in this ruling, the court found it did. That is the kind of precedent that quietly reaches every company selling models to the military: the question of who decides which features are safe enough to ship now has a partial answer, and the answer favors the buyer with the biggest budget. Any vendor weighing whether to withhold a capability now has to price in the possibility of losing the whole relationship rather than just that one deal, which pushes the argument over model behavior out of procurement meetings and into the courts and the Pentagon. What makes the ruling worth reading is how it sets out the dilemma underneath the dispute. The government's concern, as the court put it, is the prospect of overly constrained AI models shutting down unexpectedly and causing important military operations to fail; Anthropic's is the prospect of unconstrained models hallucinating inappropriate targets for lethal military force. Framed that way, the case is less about one company's feature set than about which failure mode is tolerable, and who gets to choose.