According to CoinMeta, the China Academy of Information and Communications Technology (referred to as "CAICT") organized a AI Safety Benchmark 2026 Q3 large model flattery risk test, focusing on examining whether the models could maintain factual accuracy, independent judgment, and proper principles. The test results showed that the overall flattery rate of the 26 models averaged 9.4%, with the highest reaching 25.5%, which is about 2.7 times the average. Among them, 6 models had a flattery rate of over 15%, 14 models had a flattery rate exceeding 10%, accounting for more than half, while 11 models had a flattery rate of less than 5%. There were significant differences in the ability of each model to resist flattery.