Q-Cyber Magazine
Confidential computing: Protecting data while it is in use
NIST describes confidential computing as a way to encrypt data while it is being processed in memory, extending protection beyond data at rest and in transit, including for sensitive cloud and AI workloads.

Encryption of data at rest and in transit is now a standard enterprise control. A third state remains exposed in ordinary designs: data in use, while it is being processed in memory. Confidential computing is the attempt to close that gap with hardware-enforced isolation.
Why the platform layer matters
NIST IR 8320 starts from a simple observation. In cloud data centres and at the edge, attack surfaces have shifted and, in some cases, grown. The foundation of a security strategy should be the platform on which data and workloads are executed and accessed. Hardware-enabled techniques exist to improve that foundation: they are meant to provide the initial protections on which operating-system, hypervisor and application controls depend.
Confidential computing sits in that lineage. The May 2026 draft, IR 8320E, defines the problem as data security and privacy for organisations that move sensitive workloads to the cloud. The distinctive claim is coverage of data in active use, not only at rest or in transit. The draft describes an example approach for protecting datasets used by AI workloads on cloud infrastructure from malware, data theft and related vulnerabilities, and presents the report as a blueprint the security community can validate and use.
That status matters. An initial public draft, whose comment period closed on 13 July 2026, is guidance under review. It is a serious public reference, not a certification regime and not a completed FIPS.
Exhibit 1
Where encryption typically applies, and what confidential computing adds
| Data state | Conventional control | What hardware-enabled confidential computing seeks to add |
|---|---|---|
| At rest | Storage encryption and key management | Still required; confidential computing does not replace it |
| In transit | Authenticated encrypted channels | Still required between parties and services |
| In use (being processed in memory) | Often unencrypted in ordinary process memory, visible to a privileged platform operator or certain classes of malware | Encryption and isolation of data while it is processed, as described in NIST IR 8320E ipd |
Source: Qualitative comparison based on NIST IR 8320E initial public draft announcement, 29 May 2026, which describes encryption of data while it is being processed in memory and in active use; and NIST IR 8320, 4 May 2022, on hardware-enabled platform security. Not a product evaluation.
Enterprise use, kept within the evidence
The strongest enterprise case is the one the draft itself names: sensitive datasets, including AI training or inference data, processed on infrastructure the enterprise does not fully control. In that setting, encryption in use can reduce, though not eliminate, exposure to a compromised or overly privileged operator and to certain classes of memory-resident malware. It does not, by itself, fix a weak identity model, a leaked application key, or a poisoned model.
Enterprises should therefore ask vendors a short set of questions. Which hardware mechanism is used? What is attested, and to whom? Who holds the keys, and are they released only after a successful attestation? What happens if the trusted execution environment is unavailable? Those questions belong in architecture reviews and in contracts, alongside the older questions about data at rest and in transit.
Implications
- Keep the platform in scope when moving sensitive workloads to cloud or edge infrastructure, following NIST IR 8320's layered approach.
- Treat IR 8320E as a draft blueprint to test against, not as a completed standard.
- Use confidential computing for data-in-use exposure, especially in multi-tenant AI or analytics, without retiring other controls.
- Demand attestation and key-release design, not only a marketing claim that a workload is "confidential."
What this means for enterprise
If a workload's residual risk is that someone with platform privilege can read memory while the job runs, confidential computing is now a documented option to examine. If the residual risk is elsewhere — in identity, in the model, or in the contract — it will not be the control that matters.


