To evaluate the dynamic risks of industrial cyber-physical saystems (ICPS) under cyber attacks, this study investigates a Markov-improved evolutionary game model. Based on the vulnerability nodes within the ICPS, a system attack-defense state transition diagram from the information domain to the physical domain is designed, providing a foundation for the Markov-improved evolutionary game analysis. First, in the single-stage attack-defense process, an evolutionary game model incorporating a parameter mechanism is studied to determine the payoffs of attack and defense entities with varying degrees of rationality and exploration after the game. Second, in the multi-stage attack-defense process, based on the single-stage attack-defense game model, transition probabilities and discount factors are introduced. The attack payoffs of different vulnerability nodes are calculated according to the attack-defense state transition diagram, enabling dynamic deduction of multi-stage attack-defense confrontations. Finally, the dynamic risks of ICPS are assessed based on the magnitude of attack payoffs. This study conducts numerical experiments and simulations of an industrial cyber-physical system model, using a boiling water power plant as the simulation object. The Markov-improved evolutionary game evaluation method is simulated using Matlab, and the dynamic risks of ICPS are evaluated based on the attack payoffs. The results demonstrate that the proposed model emphasizes the differences between the attack and defense sides, reasonably calculates the attacker’s payoffs in ICPS based on the varying levels of rationality and exploration of both parties. This provides a theoretical foundation for the dynamic risk assessment of ICPS under cyber attacks and offers significant reference value for enhancing the security of industrial cyber-physical systems.