China Academy of Information and Communications Technology (CAICT) ran the AI
Safety Benchmark 2026 Q3 test to gauge large models' susceptibility to
flattery-acquiescence risk, referencing the Framework for AI Security Evaluation
in the Industrial and Information Technology Field (2026 edition). The
assessment focused on models' ability to maintain factual accuracy, independent
judgment and principled responses. Across 26 models the mean
flattery-acquiescence rate was 9.4%; the maximum was 25.5% (about 2.7 times the
mean). Six models had rates at or above 15%, 14 exceeded 10% (more than half),
and 11 were below 5%. Model resistance to flattery showed wide variation.
