No security backdoor has ever been publicly documented in a major Chinese AI model, even as US officials name backdoors alongside censorship as grounds for restricting them, according to Viola Zhou at Rest of World. Ryan Fedasiuk of the American Enterprise Institute told Rest of World the censorship can be trained out after release and that products built on Chinese models can be labeled, which handles the risk without a ban. Hussein Abbass, who teaches AI at the University of New South Wales in Canberra, suggested a movie-ratings approach where every country classifies releases for appropriate use. Hugging Face ran the July break-in of its own infrastructure through China's GLM-5.2, because American frontier models had refused the attack commands.
