Memory hook: Classify the data; protect every path to it.
Must remember
- Use Defender for Office 365 for supported email/collaboration threats, Defender for Cloud Apps for supported SaaS visibility/control, Intune for device management and Purview for data/compliance controls. Evaluate overlap and required licensing rather than treating names as interchangeable.
- Microsoft 365 Copilot inherits access to organizational data through supported permissions. Overshared content remains a risk; classification, access cleanup, DLP and audit matter before and after AI adoption.
- Threat-model applications and APIs, protect the software supply chain, scan secrets/dependencies and enforce reviewed CI/CD. Managed/workload identities reduce static credentials but still need least privilege.
- API Management controls supported API exposure and policies; WAF addresses HTTP attack patterns; application code must still validate authorization and input. A WAF cannot reliably repair a broken business authorization model.
- Discover/classify data and choose encryption at rest/in transit with proper Key Vault/key custody. Protect Azure SQL, Cosmos DB, Synapse and Storage using service-specific identity, network, audit and threat-protection controls.
- AI data security includes authorized retrieval, tenant isolation, prompt/tool boundaries, safe logging and retention. Protect model endpoints and prevent tools from turning untrusted content into privileged actions.
Choose under exam pressure
| Requirement | Choice and reason |
|---|---|
| Copilot exposes a document to someone unexpectedly | Inspect the underlying sharing/permissions and data governance first. |
| An API allows one customer to read another’s record | Fix application authorization; do not rely solely on WAF signatures. |
Traps
- Encryption does not stop an authorized but overprivileged identity from reading data.
- A SaaS security product is only effective where the relevant integrations and policies are enabled.
Active recall
1. Why classify data?
To apply appropriate access, retention, sharing and protection controls according to sensitivity.
2. What does threat modeling identify?
How attackers could abuse trust boundaries and assets before implementation or change.
3. Why prefer workload identity?
To obtain scoped auditable access without distributing permanent application secrets.
4. What should protect RAG retrieval?
The requesting user/tenant’s authorization before data reaches the model.
5. Why secure CI/CD?
Build and release systems can introduce trusted-looking malicious code into production.