Modern computing environments, such as embedded systems and distributed learning environments, exhibit stochastic and dynamic behavior due to effects such as dynamic voltage and frequency scaling (DVFS), heterogeneous multicore architectures, thermal management, workload interference, and privacy-preserving operations such as differential privacy. All of these effects render traditional constant-parameter covert communication unstable at conventional modulation rates, resulting in impractically high bit error rates (BER) and detectability. This paper describes a unified Adaptive Multi-Resource Covert Communication Framework that uses real-time system-state monitoring and closed-loop control to tune bit duration, modulation amplitude, spreading factor, and error-correction rate. An information-theoretic model of mutual information, state-dependent SNR, channel capacity, and KL-divergence-based stealth metrics captures the statistical properties of resource-heterogeneous systems under stochastic workloads. The framework is evaluated on two important scenarios: embedded multicore systems under DVFS and thermal variation, and federated learning environments where gradient noise induced by differential privacy limits channel capacity. Evaluation results show that adaptive multi-resource covert channels deliver reliable transmission, high-throughput stability, and statistical stealth, surpassing both fixed-load and single-resource covert channels.