Channel state information (CSI) is essential for beamforming and precoding in sixth-generation (6G) multi-input multi-output (MIMO) wireless communication systems. Conventional codebook-based feedback compression methods often rely on simplified assumptions, which can limit CSI reconstruction quality, and iterative approaches have slow reconstruction. Machine learning (ML)-based autoencoder can improve CSI reconstruction quality. However, their effectiveness depends on strict compatibility between the encoder and decoder. A decoder trained with one encoder may not correctly interpret latent representations generated by another, leading to potential reconstruction failures. To address this challenge, this work introduces a Pairing ID as a verifiable lifecycle binding mechanism to ensure encoder–decoder compatibility. Furthermore, it investigates Pairing ID confusion threats arising from telecom deployment environments and ML model lifecycle management. The study also studies corresponding defense mechanisms to enhance the security of real-time CSI reconstruction systems.