TL;DR: Active inference leaves sensitive data and model weights exposed in memory, Fortanix says, so enterprise AI often stalls at the point of use. Inference runtimes then need hardware isolation and attestation to make key release and governance dependable.
Editorial analysis by NHI Mgmt Group, based on content published by Fortanix: “Confidential AI: The Security Foundation for Enterprise AI”.
Key questions
Q: Why do confidential AI controls matter more than storage encryption for enterprise AI?
A: Storage encryption protects data before and after use, but inference requires the data to exist in memory while the model runs.
Q: What breaks when AI workloads run on infrastructure that cannot be attested?
A: The trust model breaks because the organisation has no cryptographic proof that the runtime is authentic and unmodified.
A: Security teams should treat AI data as a production dependency, not a side asset.
Practitioner guidance
- Define in-use data protection requirements Classify which AI workloads handle sensitive records, proprietary models, or regulated data and require runtime protection before deployment decisions are made.
- Require attestation before key release Make cryptographic attestation a prerequisite for decrypting data or loading model weights into production inference environments.
- Extend trust checks to GPU execution paths Validate that the protection model covers the GPU pipeline, not only the CPU enclave, when the workload is inference-heavy.
Bottom line: Confidential AI addresses the security gap created when sensitive data and model weights must be decrypted during active computation.
Glossary terms: Confidential AI, Trusted Execution Environment, Remote Attestation, In-Use Data Exposure
Explore further
View Full Forum → | NHI Foundation Course → | Our Services → | Read the full analysis →
Confidential AI is best understood as a runtime governance control, not a compliance accessory. The article’s core point is that the real security break happens when data or model weights are decrypted for computation and the infrastructure layer can still see them. That changes the control objective from protecting static assets to protecting the active trust boundary. For security teams, this is a shift from storage governance to execution governance.
A few things that frame the scale:
- Companies using AI governance put over 12 times more AI projects into production, according to a Databricks report.
A question worth separating out:
Q: Should teams prioritise confidential AI before expanding agentic AI in regulated environments?
A: Yes, when agentic systems will handle sensitive records, proprietary models, or continuous internal access. Agentic workflows increase the number of actions, tools, and persistence points that must be trusted. If the runtime cannot be verified and isolated, scaling agentic AI expands exposure faster than governance can absorb it.
👉 Read our full editorial: Confidential AI closes the in-use gap for regulated enterprise AI