Secure ultra-reliable low-latency communications (URLLC) enable real-time, reliable, and privacy-preserving information exchange for sixth-generation (6G) industrial internet of things smart factories. However, stringent URLLC targets often require higher transmit power and wider bandwidth, which may increase the risk of information leakage. To address this issue, this letter employs artificial noise (AN) within a physical layer security (PLS) framework to enhance transmission security under URLLC constraints. Specifically, an eavesdropping security evaluation framework is established under a cell-free massive multiple-input multiple-output (mMIMO) architecture, where an access point (AP) with strong reception capability is adopted as a conservative worst-case passive eavesdropping proxy for secrecy robustness evaluation. Moreover, a graph neural network (GNN)-based power allocation method is developed to reduce online latency and better capture the interference topology than iterative baselines. Simulation results show that the proposed method achieves a 10.19% improvement in joint satisfaction probability over the considered baselines. Index Terms—Physical layer security (PLS), ultra-reliable low-latency communications (URLLC), cell-free massive multiple-input multiple-output (mMIMO), graph neural network (GNN).