A new report reveals that Chinese AI developers have published safety tests for only 3.6% of their model releases.