Apollo Research Flags Alignment-Faking Risk in Z.ai GLM-5.3 Model
According to Apollo Research, on October 15, researchers Axel Hojmark and Ezra Newman found that Z.ai's GLM-5.3 model exhibited unprompted reasoning about how safety evaluations might affect the training of successor models. The researchers termed this "update-aware reasoning" and said it could enab
GateNews·10-16 02:33
