Four security patterns for AI, and how to pick yours.
Before you choose a model, decide what each model may see. Four security patterns for AI, from Open to Sovereign, and how to pick yours.
Before you choose a model, decide what each model may see. The architecture follows.
Start with what each model may see.
Most companies ask which model to use first. A more useful first question is what each model should be allowed to see. The answer shapes the architecture.
We use four patterns on a security dial. Each is a starting point for a deployment conversation. Model routes, hosting and controls are confirmed for each deployment and written into its Security Level Schedule.
1. Open
Public frontier models under enterprise terms, with information controls set for the deployment. This can fit work without special processing requirements. The allowed data, redaction, access rules and request records are agreed before use.
2. Hybrid
Sensitive work uses private models; other work may use public models by separate agreement. Routing rules decide which model may process each task. The actual routes and controls are recorded for the deployment.
3. Private
Private models only. Review where records are stored and where inference happens; those are separate boundaries. Hosting and isolation details are On request until the deployment truth table is signed.
4. Sovereign
Private models on hardware the customer controls are the intended pattern. Customer-held keys and hosting choices must be confirmed in the Security Level Schedule. Availability is On request.
Start with the promises your data carries.
Choose the level that fits your obligations and operating environment. Your counsel decides which promises apply. We record the selected model routes, hosting and controls in the deployment agreement.
Data residency describes where records are stored. Inference residency describes where a model processes them. Check both before selecting a route.
These are common patterns, not a certification or security attestation. Hosting options and isolation details remain On request until confirmed for a deployment.
ASG does not give legal advice; your counsel decides which obligations apply.
Discussion
No comments yet