Enforcing approved model versions when Azure OpenAI auto-upgrades a deployment
To enforce approved model versions on Azure OpenAI, pin each production deployment to the assessed version or choose the most restrictive upgrade setting. Block unapproved configurations with Azure Policy and infrastructure-as-code, and limit who can change deployments. Compare the served version to the approved version continuously. Then add a runtime check on individual requests that allows, blocks or escalates and records evidence. Configuration controls govern what can be deployed. They do not show which version handled a given request.
A practical implementation sequence
- 1
Record the approval
For each production deployment, store the assessed model name, version, approver and date in a system of record.
- 2
Set the upgrade behavior
Choose the most restrictive upgrade setting your retirement plan can tolerate, and pin the version explicitly.
- 3
Encode it in IaC
Declare version and upgrade setting in Bicep, ARM or Terraform so changes need review.
- 4
Enforce with Azure Policy
Assign policy that denies or audits deployments outside the approved list, scoped to the relevant subscriptions.
- 5
Restrict change rights
Reserve deployment write access for the pipeline identity and a small break-glass group.
- 6
Alert on change events
Route deployment write operations to your monitoring and ticketing so every change maps to an approval.
Native controls and where each one stops
Version pinning and restrictive upgrade setting
Prevents automatic change on the deployment. Stops at the retirement date, and does not prove which version handled a request.
Azure Policy on deployment properties
Rejects or flags deployments whose version or upgrade setting falls outside your approved list. Governs configuration at create or update time, not individual requests.
Bicep, ARM or Terraform
Makes the approved version and upgrade setting reviewable in a pull request. Drift appears only if changes made outside the pipeline are detected and reconciled.
RBAC on deployment changes
Limits who can alter version or upgrade settings. Does not cover platform-initiated changes or a mis-scoped role.
Activity Log and diagnostic settings
Record management operations and can feed alerts. Detective only, and the evidence is about configuration, not each interaction.
Version drift is a change-control problem
An Azure OpenAI deployment binds a model name and a model version to an endpoint your applications call. When a deployment is configured to follow a default version, the model serving production traffic can change without a corresponding change review, approval or revalidation. The version your risk team assessed and the version answering requests then diverge.
The platform exposes this as an upgrade setting on the deployment. In broad terms the choices are to follow the default version automatically, to upgrade only when the current version expires, or to not upgrade automatically. Each choice moves the change decision between the provider and your own process. Pinning also has an end date set by the provider's retirement schedule, so it defers a change rather than removing it.
The operational reason this matters is stability. Microsoft has told NIST that model updates lead to prompt instability and unexpected performance variations, and that teams using pre-trained generative AI models must evaluate performance for their specific use. A version change is therefore a material change to your system, and it should pass through the same gate as any other.
Detecting when the served version differs from the approved version
Preventive controls can be bypassed by a misconfigured scope, a manual portal change or a setting you did not know existed. Detection closes that gap. Run a scheduled comparison between your approval record and the live deployment metadata, and treat any mismatch as an incident that opens a change ticket.
A second, stronger check uses the model version reported in the API response or deployment metadata, where it is exposed. Capture it alongside request identifiers and compare it to the allowlist for that deployment. Test this in a non-production environment first. Confirm that the reported value is granular enough to match your approved list, and observe what happens to requests in flight when a version changes. Do not assume a clean cutover.
Be clear about what detection gives you. It tells you drift happened and roughly when. It does not undo the interactions that were already served by the unapproved version, which is why the next layer matters in regulated settings.
Runtime enforcement in the execution path
Configuration controls and detection operate on the deployment. A runtime control operates on the request. If a policy engine sits in the execution path between your applications and the model endpoint, it can evaluate each interaction against an approved-version allowlist and then allow it, block it or escalate it for human approval. The decision is enforced at runtime, not reviewed after the fact.
Trussed enforces policy in the execution path in this way, at the level of individual interactions and tool calls. For version control, that means a request aimed at a deployment whose version attribute is not on the allowlist can be stopped before it reaches production users or downstream systems. The check depends on the version attribute being available to the path, so validate that early in your design.
Each decision produces audit-ready evidence: what policy applied, what the request was, and whether it was allowed, blocked or escalated. That gives a reviewer a per-interaction record that the approved version was in use. Trussed does not establish model provenance or lineage. It enforces policy on the interaction and records the outcome. Evidence is mapped to frameworks including NIST and the EU AI Act, and the platform can be self-managed in the customer cloud with configurable retention and data residency.
Revalidation windows and framework alignment
Enforcement tells you a version is blocked. Change management tells you how a new one becomes approved. Define an intake path for a candidate version, an evaluation period, a decision and an update to the allowlist. Staged release is a useful pattern: route a limited share of non-critical traffic to the candidate before broad approval.
External expectations give useful reference points. Clause language in a GSA acquisition document dated June 2026 treats model version changes as material changes. It describes written notice at least 30 calendar days before a planned change, and minimum evaluation periods of 30 calendar days for major versions and 15 for minor versions. It also describes notice within 7 calendar days of identifying a change that materially increases output bias, decreases safety guardrails or behavioral constraints, or degrades performance or truthfulness. Whether such terms apply to you depends on your contracts, so use them as a model for your own windows rather than a universal rule.
The NIST AI RMF 1.0 and the Generative AI Profile are voluntary frameworks that support the same discipline. The Generative AI Profile advises documenting trade-offs, decision processes and measurement results, and considering staged release approaches. Your approval record, revalidation results and runtime evidence together form that documentation.
Common questions
Short answers to questions platform teams raise when scoping this work.
- {"question":"Is pinning a version enough?","answer":"Pinning prevents automatic change on the deployment, but it does not prove which version served a given request, and it ends when the version retires. Pair it with policy, drift detection and a per-request check where your risk profile requires evidence."}
- {"question":"Can Azure Policy alone enforce approved versions?","answer":"Azure Policy can deny or audit deployment configuration outside an approved list. It acts on resource properties, not individual requests, so it cannot show which version handled a specific interaction."}
- {"question":"What evidence should an auditor expect?","answer":"A recorded approval per deployment, change events mapped to tickets, revalidation results, and a per-interaction record of the policy decision. Together they show only approved versions served production traffic."}
Layers of version control
Label: Deployment configuration
Description: Pinned version and restrictive upgrade setting
Label: Policy-as-code
Description: Azure Policy and IaC reject unapproved settings
Label: Drift detection
Description: Compare served version to the approved record
Label: Runtime enforcement
Description: Allow, block or escalate each request, with evidence
Add a request-level check to your version controls
See how policy enforcement in the execution path allows, blocks or escalates individual interactions and produces audit-ready evidence for a regulated enterprise.
Request a DemoHow does your AI governance program compare?
See where your program has gaps in less than 2 minutes.