Architecture-specific review

Azure OpenAI security review

A secure model endpoint does not establish that the application around it is safe to ship. Review who can invoke the model, which data enters the prompt, what the answer can influence, and the controls demonstrated by the implementation.

Request access →

Review scope

What Drel establishes for this stack

Scope the actual application and its deployment. A private endpoint or a content filter is a useful control, but neither proves that retrieved documents are authorized, agent actions are bounded, or a human can defend the decision.

Drel rebuilds the system from its repository, maps the attack paths through each area, checks the controls it can verify in your code, and proposes the fix for each control that fails. A reviewer decides what counts as evidence.

01

Identity and invocation

Which workload and users can invoke the Azure OpenAI resource? Where are permissions checked before a user request becomes a model request?

What counts as evidence: Entra role assignments, workload identity configuration, application authorization logic and the route from originating user to model call.

02

Network and data boundary

Is access to the resource restricted as intended, and where can prompts, responses and retrieved content travel?

What counts as evidence: Resource network configuration, private endpoint and DNS design where applicable, data-flow diagram, logging destinations and retention settings.

03

Retrieval and prompt assembly

Can a document from a lower-trust source inject instructions or expose content the requester should not see?

What counts as evidence: Ingestion and retrieval design, source labels, access-filter enforcement, prompt assembly code and your own tests with untrusted documents.

04

Tools and consequential actions

Can the assistant call a downstream API or change a business record? Whose authority does that call use, and which actions require approval?

What counts as evidence: Tool definitions, server-side authorization, delegated identity or service permissions, approval gate and audit trail.

05

Response handling and decision

Which risks are covered by content filters and application checks, which remain open, and who signs off the residual risk?

What counts as evidence: Deployed filter policy, application tests, open control gaps, accountable owners, review conditions and a dated review record.

Architecture example

Follow authority and data end to end

A model endpoint is one component. The security decision also depends on the identity that calls it, the data it receives, the actions it can trigger, and who accepts the remaining risk.

For an internal support copilot: employee identity → application authorization → retrieved ticket and knowledge base content → Azure OpenAI → suggested response or ticket action. Review the access check at retrieval and again before any ticket update; the model response is never the authorization decision.

From repository to approval

Drel models the system from its repository, threat-models every attack path against OWASP, the EU AI Act and 22 more, maps the attack paths through it, checks the controls it can verify in your code, proposes the fix in a Jira ticket, and re-checks the automatic checks you turn on when the branch you link changes. Your AI Committee signs the approval. Unknown or missing evidence remains visible; it is not treated as a passed control.

Read the review methodology →

Further reading