“Companies like Palantir and Scale AI have advertised large language model-based systems that mix sensible military uses with highly problematic ones, such as generating "non-escalatory" courses of action.”