U.S. AI companies and the government are increasingly concerned about the effectiveness of international competition using distillation attacks to glean valuable data from frontier models to train cheaper, faster alternatives overseas.