Microsoft SC-500: Data Security Posture for AI
AI security posture is not a list of model endpoints. It is the evidence connecting each model, data source, agent identity, network boundary and workload deployment to the risks an attacker could exploit. The useful unit of review is the complete application: its retrieval store, delegated tools, credentials, deployment pipeline and logging systems. A secure model endpoint cannot compensate for an overprivileged agent or a search index that returns data to the wrong audience.
On this page
- Inventory the complete AI application before assessing posture
- Connect AI code, runtime and data
- Differentiate AI resource visibility from agent-level visibility
- Prioritize reachable attack paths, not isolated finding volume
- Protect retrieval data and derived AI data separately
- Reduce the authority available to agent tools
- Connect private networking with service-level authorization
- Treat telemetry, prompts and traces as protected data
- Turn posture findings into ownership and prevention
Inventory the complete AI application before assessing posture
Begin with an application diagram that includes the user or calling service, model deployment, retrieval system, storage, database, search index, API tools, network path and operational log stores. A model can be well configured while the data it retrieves is overexposed or the agent’s service principal can modify resources unrelated to its job. Posture should describe the actual system, not just the AI resource shown in a portal.
Microsoft Defender Cloud Security Posture Management can discover supported AI workloads and form an AI bill of materials connecting applications, libraries, data and supporting resources. The value of that inventory is its context. Security teams should know which development repository produced a deployed component, who owns the workload, what environment it serves and whether the component can reach sensitive or regulated information. If those relationships are missing, remediation will tend to become a list of unassigned warnings.
Use discovery as a starting point rather than a completeness guarantee. An AI application can call a third-party API or expose a custom tool that one cloud inventory never sees. Reconcile Defender findings against the team’s architecture records and application inventory. Assign one person to resolve unknown or ownerless assets before classifying their findings as low priority.
Use the AI bill of materials to connect code, runtime and data
An AI bill of materials should help an engineer trace a problematic library or endpoint back to the application and data it affects. The chain may include a container or dependency, deployment identity, Azure AI resource, model version and retrieval datastore. The risk of a library vulnerability depends on whether the vulnerable code executes in an internet-facing service or an isolated build environment and which secrets, identities and datasets it could reach.
Defender for Cloud can contribute source-code and infrastructure-as-code misconfiguration signals alongside cloud-resource posture. Those signals become more actionable when the team records deployment versions and keeps a secure pipeline. A manual fix to one running AI resource may be erased by the next automated deployment if the insecure setting remains in Bicep, Terraform or the application template.
When triaging one issue, ask which upstream artifact or template produced it and where the same pattern appears elsewhere. For example, correcting public access for a single model service is a short-term fix; updating a shared private-endpoint template can prevent repeated exposure across later deployments. That distinction is how AI posture becomes an engineering control rather than a dashboard.
Differentiate AI resource visibility from agent-level visibility
Microsoft’s current documentation distinguishes AI resource discovery from agent-level inventory. Defender CSPM continues to discover supported Microsoft Foundry accounts and projects; the agent-specific discovery and security-posture capabilities for Foundry and third-party cloud agents are currently preview and, since July 1, 2026, require Microsoft Agent 365 licensing. Organizations should not assume that seeing a Foundry resource means they are licensed to assess every agent acting within it.
The practical approach is to document which tenant and subscription contain the workload, which plans are enabled, which capabilities are generally available, and which features or API responses are still preview. This prevents a security review from asserting coverage the organization does not own. The Microsoft AI security posture guidance should be checked again before production rollout because licensing and scope are time-sensitive.
Where agent inventory is incomplete, application teams still need their own registry. Record each agent’s purpose, owner, execution identity, permitted tools, user delegation pattern, data sources and removal procedure. An unowned tool integration can create a larger breach path than the underlying model, even when the cloud resource has an otherwise clean configuration.
Prioritize reachable attack paths, not isolated finding volume
A public model endpoint, an overprivileged workload identity and a private dataset may look like three unrelated configuration findings. Together they can form a path from an attacker-controlled request to sensitive data. Defender CSPM’s attack-path analysis connects resource exposure, permission relationships and vulnerable configurations to help identify paths with meaningful consequences.
Start remediation by asking which control breaks the high-consequence path with the least disruption. The best answer may be to disable public access, constrain a service principal, patch a vulnerable dependency or remove a tool’s broad write permission. Track the business importance of the data and the plausibility of the path rather than treating every recommendation’s severity as independent truth.
Validate whether the path is actually broken after the change. An endpoint might still be accessible through a secondary gateway, or a user-assigned managed identity might retain equivalent permissions on another resource group. Use repeatable authorization and network tests as evidence that risk decreased. The related Defender attack-path analysis article focuses on this path-breaking design process.
Protect retrieval data and derived AI data separately
A retrieval-augmented generation system often copies information into a search index, vector representation, prompt context, response, trace or long-lived agent memory. The original source may enforce detailed user permissions while a shared retrieval index exposes content through a more permissive application identity. Posture review must map those derived copies and verify where source authorization is preserved or re-applied.
For each dataset, record classification, tenant ownership, caller identity, intended use, retention and deletion behavior. Encrypt and restrict access where the underlying platform supports it, but do not confuse encryption with authorization. A private storage account that can be read by a broad runtime identity remains exposed to a compromised agent or API.
A security test should ask one user or agent without permission to the source document to retrieve a related answer through the AI application. If the model returns restricted content, the problem may be in retrieval authorization rather than the model service’s network settings. Treat the result as a data-control failure and follow it through the index and tool chain.
Reduce the authority available to agent tools
An agent with a read-only search tool and one with an administrative deployment tool do not have the same security consequence. Enumerate tools, calling identities, delegated user context, approved actions and any approval checkpoints. The tool that actually changes a business record should perform deterministic authorization; model instructions alone cannot prove an operation is safe.
Managed identities remove static credentials but do not remove authorization work. A service principal with Contributor across a subscription can still cause serious damage if the compute running it is compromised. Give tools narrowly scoped roles and distinct identities where practical, and alert on permission expansions that materially change blast radius.
Threat-model a malicious retrieved instruction that asks the agent to export a confidential report or invoke an unrelated administrative action. The expected result is a denied operation with enough audit evidence to explain why the request was blocked. For broader execution risks, see cloud and AI threat modeling.
Connect private networking with service-level authorization
Private Link, service firewalls and network segmentation can reduce an AI application’s reachable surface. They do not prove that the model, retrieval service or database is authorized to serve a specific caller. A secure design should trace both the network path and the token or credential used at each service boundary.
Check that private DNS resolves correctly from intended workloads, that the public service endpoint is disabled when private-only access is required, and that a different workload cannot use an overly broad role assignment to reach the same data. A common failure occurs when a team adds a private endpoint but forgets to update an existing public firewall exception created during the pilot.
During audit, record both successful and prohibited connection results and the associated caller identity. A 403 may indicate authorization, while a timeout or incorrect IP can indicate DNS or routing. Avoid broadening the network rule before identifying which layer actually failed. The Private Link security patterns discussion covers endpoint placement, DNS and failure recovery.
Treat telemetry, prompts and traces as protected data
AI observability can capture user prompts, retrieved snippets, model output, tool arguments and exception data. These records may be as sensitive as the underlying business database. Apply role restrictions, retention and redaction appropriate to the use case, and avoid logging whole confidential source documents merely to make debugging easier.
Create stable correlation IDs across identity sign-in, application request, retrieval, model invocation and tool action so responders can reconstruct what happened. Define which signals come from Defender CSPM, which come from workload-protection services and which depend on the application’s own tracing. A clean posture score does not establish that misuse of a write-capable tool will be logged in enough detail for investigation.
Test one approved and one denied tool call, then confirm the operator can find the relevant evidence without being granted unrestricted access to unrelated users’ conversations. Retention must reflect privacy, legal and operational requirements, including deletion when information is no longer needed. An auditor should be able to explain both why telemetry exists and who may inspect it.
Turn posture findings into ownership and prevention
A workable AI posture program assigns each risk to the team that can change the relevant deployment artifact, identity, network path or dataset permission. Engineers should prioritize risks by reachable exposure, sensitive data, privilege and actual workload importance. The outcome metric is an eliminated attack path or corrected insecure pattern, not simply a higher security score.
Review the system whenever an AI workload adds a data source, external model provider, new tool, persistent memory or cross-tenant identity. These changes can create material security exposure without changing the name of the original Azure resource. Trigger reassessment from capability changes, not only from scheduled audit dates.
For Microsoft SC-500 preparation, a realistic final lab is to deploy a modest model-plus-retrieval application, constrain its identity, configure the intended private service paths, review Defender for Cloud posture, simulate one unauthorized data request and show the relevant logs. The engineer who can explain which controls prevented the misuse has learned more than the engineer who memorized one AI BOM dashboard.