CoreWeave Launches 'Zero Trust AI Framework' Integrating Key Control and Confidential Computing for Enterprise Security

US-based AI cloud infrastructure company CoreWeave announced the "CoreWeave Zero Trust AI Framework™" on October 21, 2026. The framework aims to address the core trust issue enterprises face when migrating sensitive data to the cloud for AI training and inference: whether the cloud provider retains access to customer data and models. CoreWeave states that the framework shifts security responsibility from "relying on vendor policies" to "technical boundaries controlled by the customer" through six layers of protection.

Previously, CoreWeave launched Remote Key Encryption (RKE) in September, allowing customers to retain encryption keys in their own key management systems, ensuring CoreWeave holds only ciphertext. This new framework expands on that foundation with additional dimensions, including Support Access Management (SAM), hardware root-of-trust attestation, and planned confidential computing support for NVIDIA Blackwell and Vera Rubin platforms. Notably, SAM allows customers to decide when CoreWeave engineers are permitted to enter the infrastructure for troubleshooting, breaking the traditional cloud service model where providers have default privileged access.

It is important to note that some components remain in planning or early stages. In its blog post, CoreWeave explicitly uses the phrase "intends to support" to describe its confidential computing capabilities on NVIDIA Blackwell and Vera Rubin platforms, indicating that while the overall architecture is announced, certain advanced security features are not yet fully available across all environments. Dave McCarthy, Group Vice President at IDC, commented that this move represents a critical step in shifting trust from vendor promises to technical controls, which is essential for enterprises deploying proprietary AI at scale.