When an underwater vehicle (UV) suffers damage and flooding, the flooding process continuously disrupts the buoyancy balance and attitude stability of the hull, increasing the risk of instability and depth loss. Dynamic anti-sinking is a crucial emergency response measure after a UV suffers damage and flooding. Traditional methods based on rule-based reasoning, case-based reasoning, and even expert systems have limitations in adaptability problems under complex dynamic environments. This paper addresses the dynamic anti-sinking decision-making problem in underwater vehicle (UV) damage and flooding scenarios. A reinforcement learning decision model is constructed, and safety constraints are introduced within a maximum entropy reinforcement learning framework. A dynamic anti-sinking decision-making method based on the Soft Actor-Critic with Lagrangian (SAC-Lag) algorithm is proposed. Simulation analysis of three damage scenarios is conducted to evaluate the dynamic anti-sinking maneuvering indices and decision-making viability indices of the proposed method, and evaluation criteria are defined. The results show that the proposed SAC-Lag can achieve safe surfacing under light to moderate damage conditions, and exhibits a lower peak trim and higher viability score compared to the standard SAC. In severely damaged conditions, although it is difficult to achieve ascent within a limited time by relying solely on rudder control, this method improves the SAC vitality score by 11.5 and has better overall anti-sinking performance.