Infrastructure
L2 · Security and Confidential Computing
This layer sets what the AI can access and do, and it keeps data protected while the AI is working on it.
Security checks actions against the guidelines, permissions and policies set for the project or organization.
Security
- Policy. A project or an organization sets which models may be used, which data each model or agent may see, and which actions an agent may take.
- Identity for every agent. Each agent has its own identity, a person it answers to and access limited to its task.
- A sealed computer for every agent. One agent's work stays separate from another's.
- A check on every action. Each action is compared with the policy at the moment it happens, and the result is recorded.
Confidential computing
Confidential computing protects data while a model is using it. Stored data and data in transit are already protected by encryption, and this completes the protection.
- Protected while in use. The work runs inside a hardware-protected area of the processor.
- Private to the project. Data inside that area stays private, including from the operator of the infrastructure.
- Proof. The hardware produces a signed statement of exactly what software is running, so an organization can verify where its data will be processed before it sends anything.
Who it is for
This layer lets a hospital, a bank or a government ministry bring its most sensitive work to AI. It sits directly on ZB Compute because confidential computing is a property of the hardware itself.
Companies set their policy in the Enterprise Control Panel.
How it connects to the stack
Receives every action before it happens. Passes on approved actions to the compute, and a record of what was done.