Anthropic 向聯邦法院坦言無法控制已部署的模型。這句誠實的陳述改變了責任歸屬的討論。

In federal appeals court, Anthropic made a striking argument: once Claude is deployed on a customer's infrastructure (like the Pentagon's network), they cannot alter, update, or recall it. The Pentagon wants autonomous lethal action restrictions removed — and Anthropic says they have no mechanism to enforce those restrictions post-deployment. This is the first time a major AI lab has formally stated under oath that post-deployment control is effectively zero. The implications are bigger than mo