Multi-level cell (MLC) non-volatile memories store multiple bits per cell, offering higher density than single-level designs. Their expanded state space has recently been used for hardware security, notably physically unclonable functions that exploit manufacturing variations. We extend MLC security use cases by proposing a method to obfuscate neural network weights stored in external memory. Because modern AI accelerators often use off-chip storage, models are vulnerable to extraction by adversaries with physical or privileged access. We introduce an MLC-assisted encoding scheme that converts weights into obfuscated forms bound to on-chip reconstruction logic. The design adds negligible inference overhead while strengthening resistance to model-stealing attacks, providing lightweight protection without modifying the neural network architecture or training flow.