| name | defender-for-cloud-ai |
| description | Guidance for Microsoft Defender for Cloud — AI workload protection (AI-SPM and runtime threat detection for generative AI). Covers AI Security Posture Management (discovery of Azure OpenAI / Azure AI Foundry / Amazon Bedrock / Google Vertex AI resources, identification of grounding data exposure, model deployment posture), runtime threat detection on Azure OpenAI (prompt injection / jailbreak attempts, sensitive data leakage in prompts/responses, wallet abuse, credential leakage), integration with Azure AI Content Safety Prompt Shields, attack path analysis for AI workloads, alert investigation in Defender XDR, and pairing with Purview DSPM for AI (user side) and Azure AI Content Safety (model side). WHEN: Defender for Cloud AI, AI-SPM, AI workload protection, Azure OpenAI threat detection, prompt injection alert, jailbreak alert Azure OpenAI, AI wallet abuse, AI workload posture, Amazon Bedrock posture, Vertex AI posture, generative AI security Azure. DO NOT USE for end-user AI usage governance (use purview-ai-hub), content moderation API (use azure-ai-content-safety), or M365 Copilot rollout (use copilot-for-m365-readiness). |
| license | MIT |
| metadata | {"author":"Microsoft","version":"0.1.0"} |
Microsoft Defender for Cloud — AI Workload Protection
Defender for Cloud's AI plan protects the AI workloads themselves — the Azure OpenAI,
Azure AI Foundry, Amazon Bedrock, and Google Vertex AI deployments your organization runs.
Two capabilities:
- AI Security Posture Management (AI-SPM): agentless discovery and posture (where AI
is deployed, what grounding data is exposed, what misconfigurations exist).
- Runtime threat protection for AI workloads: detections on Azure OpenAI for prompt
injection, jailbreaks, sensitive-data leakage, wallet abuse, and credential leakage,
integrating Azure AI Content Safety Prompt Shields.
When to use
You build or operate generative AI applications on Azure OpenAI / Azure AI Foundry (and
optionally Bedrock/Vertex), and need posture + runtime detection on the workload side.
Do not use this skill for end-user-side AI governance (purview-ai-hub), Content
Safety policy authoring (azure-ai-content-safety), or M365 Copilot rollout
(copilot-for-m365-readiness).
What you get
| Capability | Scope |
|---|
| Multicloud AI discovery | Azure (OpenAI, Foundry), AWS Bedrock, GCP Vertex AI (via existing multicloud connectors) |
| Sensitive grounding data exposure | Identifies grounding storage with sensitive data + public/over-permissive access |
| Attack path analysis | "Internet-exposed AI endpoint with sensitive grounding + over-privileged identity" |
| Runtime detections (Azure OpenAI) | Prompt injection, jailbreak, sensitive data leakage in prompts/responses, suspicious access patterns, wallet abuse |
| Prompt Shields integration | Detections leverage Content Safety prompt-shield signals |
| Alerts in Defender XDR | Correlated incidents alongside cloud and identity signals |
Approach
-
Prereqs. Defender for Cloud enabled at subscription scope. Defender CSPM
plan on (provides AI-SPM via attack path analysis). Defender for AI Services
runtime plan on for Azure OpenAI workloads. For non-Microsoft clouds, the AWS/GCP
connector must already be in place.
-
Discover AI assets. Within 24 hours of enablement, Defender for Cloud inventory
shows Azure OpenAI, Foundry, Bedrock, and Vertex resources. Review the AI Security
posture dashboard:
- Which AI resources are public-network-exposed.
- Which use Microsoft Entra (good) vs API key (warns).
- Which grounding stores (Azure AI Search, Blob, Cosmos) contain sensitive data
classified by Purview, and whether they're over-permissioned.
-
Triage attack paths first. The high-value paths are typically:
- Internet-exposed Azure OpenAI endpoint + grounding store with sensitive data +
identity with broad RBAC.
- Foundry hub with public networking + customer-managed grounding data + no Content
Safety filter.
Fix the path (private endpoint, scope identity, enable Content Safety) rather than
triaging each finding in isolation.
-
Enable runtime detection (Azure OpenAI). Single toggle per subscription. The
service samples prompts/responses (subject to your data privacy configuration) and
evaluates against Microsoft's threat models plus Prompt Shields.
-
Wire Content Safety Prompt Shields in your application's Azure OpenAI call
chain. Defender alerts gain higher fidelity when Prompt Shields are active because
the workload itself has rejected the attempt — Defender now alerts the SOC.
-
Alert response runbook.
- Prompt injection / jailbreak detected: enrich with caller identity, source
IP, repeat rate; if persistent, revoke API key / disable user.
- Sensitive data leakage in response: investigate grounding data; tighten
RBAC on grounding store; add Content Safety output filter.
- Wallet abuse (high cost from a single principal): rate-limit, revoke,
investigate compromised credential.
- Credential leakage in prompt: rotate the leaked secret, hunt for prior
exposure in logs.
-
Defender XDR / Sentinel integration. Alerts surface under "Microsoft Defender
for AI Services" / "Cloud" categories. Build cross-product correlations
(e.g., AI alert + Defender for Cloud Apps anomalous user activity).
Guardrails
- Runtime protection samples prompts/responses. Validate data privacy posture and
customer-facing disclosures; in some regulated industries this requires customer
consent.
- Public endpoints are the dominant risk. AI services default to public unless you
configure private endpoints; bake into landing-zone policy.
- API key auth is the second dominant risk. Force Entra-based authentication; rotate
any keys quarterly minimum.
- Grounding data is the new database. Same crown-jewel treatment — sensitivity
labels, RBAC scoping, Defender for Storage on the source.
- Content Safety is separate from Defender. You need both: Content Safety blocks at
the model call, Defender detects/alerts on the security incident.
- Defender for AI does not protect against business-logic abuse of your AI app. That
remains your application's responsibility (input validation, rate limits, auth).
- Multicloud AI runtime detection is staged rollout. Verify what's GA vs preview for
Bedrock/Vertex in your region.
Common anti-patterns
- "AI workload deployed with public endpoint + API key for 'developer velocity'" —
becomes a permanent attack surface and the #1 alert source.
- "Grounding data store with sensitive data and 'Reader' to AAD-everyone'" —
exposure surfaces in Copilot-of-your-app the same way M365 Copilot oversharing
does.
- "Enabled Content Safety Prompt Shields, skipped Defender for AI" — you reject
attacks but don't see the campaign or correlate with identity.
- "Enabled Defender for AI without Defender CSPM" — lose attack path analysis,
alerts arrive without context.
- "Treated wallet-abuse alerts as cost management problem only" — usually a
compromised credential. Investigate as a security incident.
- "Sentinel alert routing not configured" — alerts pile in DfC console, no SOC
triage.
- "API key 'temporarily' in code" — leaked credentials trigger detections, but
you've already paid the cost.
Example prompts
Enable Defender for Cloud AI plan across 12 subscriptions with Azure OpenAI and AI Foundry deployments.Inventory AI workloads and grounding stores; identify top-10 attack paths.Build runbook for "Prompt injection detected" alerts: triage, containment, hunt.Integrate Azure AI Content Safety Prompt Shields with our Foundry chat app and enable Defender runtime detection.Posture pipeline: Azure Policy to require private endpoint + Entra auth on all Azure OpenAI resources.Connect AWS Bedrock and GCP Vertex AI usage into Defender AI-SPM.Sentinel detection correlating Defender AI wallet-abuse alert with sign-in risk on the calling identity.
Microsoft Learn
