Modern military operations are increasingly evolving into Network-Centric Warfare (NCW), where the security of tactical communications has become a critical concern. Traditional cryptographic methods, while ensuring confidentiality, inadvertently reveal the presence of sensitive messages, making them potential targets for detection and analysis. To address this limitation, this study proposes a novel steganographic approach that conceals the existence of tactical messages by embedding them within natural language text, thereby achieving both secrecy and plausible deniability. Leveraging recent advances in generative AI, particularly the LLaMA-2 language model, we develop a text-based steganography system tailored for military communication. A custom dataset was constructed by combining the UCDP conflict event records with the literary corpus of Treasure Island, enabling domain adaptation through fine-tuning with QLoRA. The resulting system demonstrates strong performance across multiple evaluation metrics, achieving a message reconstruction accuracy of 90.70%, an evasion rate of 98.25% in terms of ROC-AUC against detection models, a perplexity score of 25.06 indicating high linguistic naturalness, and an average embedding capacity of 4.5 bits per word. Compared to conventional cryptographic and steganographic techniques, the proposed method significantly improves information hiding capacity while reducing detectability, thus presenting a viable solution for evading enemy interception and traffic analysis in military environments. Future work will focus on optimizing the framework through reinforcement learning, extending the system to multimodal data, and developing lightweight models for real-time, high-reliability deployment in tactical communication scenarios.