“The safety guardrails of commercial large language models could hinder their use in core U.S. Department of Defense applications, as queries related to lethal operations might violate usage policies.”