We propose a novel approach to achieve physical layer security for integrated sensing and communication (ISAC) systems operating in the presence of targets that may be eavesdroppers. The system is aided by a time-modulated intelligent reflecting surface (TM-IRS), which is configured to preserve the integrity of the transmitted data at one or more legitimate communication users (CUs) while making them appear scrambled in all other directions. The TM-IRS design leverages a generative flow network (GFlowNet) framework to learn a stochastic policy that samples high-performing TM-IRS configurations from a vast discrete parameter space. Specifically, we begin by formulating the achievable sum rate for the legitimate CUs and the beampattern gain toward the target direction, based on which we construct reward functions for GFlowNets that jointly capture both communication and sensing performance. The TM-IRS design is modeled as a deterministic Markov decision process (MDP), where each terminal state corresponds to a complete configuration of TM-IRS parameters. GFlowNets, parametrized by deep neural networks are employed to learn a stochastic policy that samples TM-IRS parameter sets with probability proportional to their associated reward. Experimental results demonstrate the effectiveness of the proposed GFlowNet-based method in integrating sensing, communication and security simultaneously, and also exhibit significant sampling efficiency as compared to the exhaustive combinatorial search and enhanced robustness against the existing benchmarks of physical layer security.