Recently, deep learning of encoding and decoding functions for wireless communication has emerged as a promising research direction and gained considerable interest due to its impressive results. A specific direction in this growing field are neural network-aided techniques that work without a fixed channel model. These approaches utilize generative adversarial networks, reinforcement learning, or mutual information estimation to overcome the need of a known channel model for training. This paper focuses on the last approach and extend it to secure channel coding schemes by sampling the legitimate channel and additionally introduce security constraints for communication. This results in a mixed optimization between the mutual information estimate, the reliability of the code and its secrecy constraint. It is believed that this lays the foundation for flexible, generalizable physical layer security approaches due to its independence of specific model assumptions.