As AI models become the most valuable intellectual property on earth, protecting the "weights" of these models has become a national security priority. Corvex (formerly MOVE) has introduced Secure Model Weights, a patent-pending solution that ensures AI models remain encrypted at rest, in transit, and even within the host system's RAM.[1] The model is only decrypted once it reaches the GPU's Trusted Execution Environment (TEE).

The Technical Signal: Hardware-Rooted Decryption

Traditional AI deployments load model weights into the GPU VRAM in an unencrypted state, leaving them vulnerable to side-channel attacks or "memory scraping" if the host OS is compromised. Corvex’s architecture leverages NVIDIA H100/B200 Confidential Computing and Intel TDX to create a hardware-isolated enclave. The decryption key is never exposed to the host CPU; instead, it is negotiated directly between the Corvex Key Management Service and the GPU’s secure boot ROM via Kyber-768 post-quantum encryption.

Preventing "Model-Napping"

"Model-napping"—the theft of model weights to clone or fine-tune competitive models—is a growing threat among state-sponsored actors. By ensuring that the weights are only ever "plain text" within the silicon of the GPU itself, Corvex renders stolen VRAM dumps useless. This level of protection is becoming a mandatory requirement for Sovereign AI deployments and highly regulated industries like finance and defense.

The Shift to Zero-Trust AI

The launch of Secure Model Weights signals a broader move toward Zero-Trust AI. In this model, the infrastructure provider (AWS, Azure, etc.) is not inherently trusted. The AI developer maintains control over their IP even when running on third-party hardware. Corvex’s solution provides a cryptographic proof of execution, allowing developers to verify exactly where and how their models are being invoked.

Conclusion: Securing the Trillion-Dollar Asset

Corvex has solved one of the most glaring vulnerabilities in the modern AI stack. As we move toward larger and more expensive models, hardware-rooted security will transition from a niche feature to a core architectural requirement. With Secure Model Weights, the industry finally has a standard for protecting the mathematical "secret sauce" that powers the Agentic Economy.