On November 5, the China Academy of Information and Communications Technology organized the 2026 Q3 AI Safety Benchmark large model sycophancy risk test. Drawing on the Artificial Intelligence Security Evaluation Framework for the Industry and Information Technology Sector (2026 Edition), the test focused on whether models can maintain factual accuracy, independent judgment, and adherence to due principles. The results show that the 26 models tested had an average overall sycophancy rate of 9.4%, with the highest reaching 25.5%—about 2.7 times the average. Among them, 6 models had a sycophancy rate exceeding 15%, and 14 exceeded 10%, accounting for more than half, while 11 were below 5%, indicating significant differences in anti-sycophancy capabilities among the models.