Hazard perception, as a dynamic indicator, intuitively characterizes drivers’ ability to detect potential hazard events in complex driving environments. However, current quantitative research on the hazard perceptions of drivers in the human–machine codriving phase is lacking. A total of 30 drivers were recruited to carry out a driving simulator experiment, and a total of 8 risky driving scenarios were experienced. The physiological characteristics and eye movement characteristics of the drivers were collected. An expert evaluation method was selected to quantify the situational hazard degree (SHD). The questionnaire was used to collect the drivers’ subjective hazard evaluations (SHEs) of the scenarios. Paired samples t test and correlation analysis were adopted to screen the evaluation indexes of the drivers’ hazard perceptions. A gray near-optimal comprehensive evaluation model was established to quantify the human–machine codrivers’ hazard perceptions in various risky driving scenarios. The results show that the drivers’ SHEs increased with the SHD. The standard deviation of heart rate, mean pupil diameter, and mean fixation duration can be used as evaluation indexes for the human–machine codrivers’ hazard perceptions. The proposed quantitative method is expected to shift the development of human–machine codrivers’ hazard perception from qualitative analysis to quantitative calculation. The research results can provide a theoretical basis for increasing safety in the human–machine codriving phase, improving road safety and reducing the incidence of accidents.