Again, the statute says nothing about "dictating terms." It covers an adversary who might sabotage or subvert a system. And no, Anthropic did not cancel anything, did not pull any capabilities, and has no backdoor into deployed models. There was no "um actually." It wrote two prominent usage restrictions into a contract, the military read and signed it, and then later demanded the terms be removed.

As a sanity check, I’d highly recommend trying out your own line of reasoning in cases which you think you might feel differently about. For example, imagine it's instead a certain authoritarian state saying "obey the Party or get blacklisted as a national security threat."

It might feel good in the moment, like "yeah, Anthropic, don't tell the military what to do," because you agree with this particular call. But the government doesn't give that power back, and you might not be so happy about the next way it gets used.