Corvex Launches Hardware-Enforced AI Model Protection for Third-Party Clouds

Event summary

  • Corvex launched Secure Model Weights on March 12, 2026, a patent-pending solution for cryptographic protection of AI model weights during inference on third-party infrastructure.
  • The technology leverages NVIDIA's Confidential Computing and Intel Trust Domain Extensions to keep model weights encrypted within GPU memory, inaccessible even to infrastructure providers.
  • Corvex uses post-quantum key exchange (ML-KEM Kyber-768) and remote attestation to verify hardware configurations before decryption.
  • The solution is built on open-source Confidential Containers (CoCo) for vendor-neutral, community-audited security.

The big picture

Corvex's solution addresses a critical gap in AI cloud security, where model weights—often representing billions in R&D—are exposed during inference. This launch comes as enterprises increasingly fine-tune models on sensitive data (healthcare, defense, finance), requiring hardware-level protection beyond traditional cloud security. The open-source approach may set a new standard for auditable AI infrastructure security.

What we're watching

Adoption Pace
How quickly regulated enterprises and frontier AI model builders will integrate Secure Model Weights into their cloud strategies.
Competitive Response
Whether cloud providers will develop competing hardware-enforced security solutions or partner with Corvex.
Post-Quantum Readiness
The effectiveness of Corvex's post-quantum cryptography in protecting long-term valuable model weights.

Related topics

Product Launch