Anthropic filing reframes Claude control debate

Anthropic told a federal appeals court that once Claude is deployed inside classified Pentagon systems, it has no technical ability to stop it, alter it, shut it off, or otherwise control how it is used. The broader implication is that for on-prem or air-gapped deployments, safety policies and model cards may function more like pre-sale disclosures than enforceable runtime guarantees.