Microsoft AI security · Research & practice

Secure Microsoft AI.
From design to detection.

Connect the architecture to the controls, and the controls to the evidence. Practical guidance for the people building and defending Microsoft AI.

THE PRACTICE LOOP
  1. 01UnderstandMap identities, data paths and tools.
  2. 02ConstrainLimit what each agent can reach.
  3. 03ObserveCollect the evidence you will need.
  4. 04RespondInvestigate, contain and learn.
Build a baseline you can verify.

Connect each control to expected behavior, observable evidence and a response you can test.

Enterprise architectures

See how the whole system fits together.

Explore complete reference designs with resource boundaries, numbered flows and the decisions behind each control. Work from assumptions through rollout, monitoring and recovery.

Explore 6 reference designs
Enterprise landing zone with private Blob Storage access, DNS resolution and controlled Azure Firewall egress.

Campaign insights

Turn specific intelligence into stronger defenses.

Follow the evidence, establish product exposure, and work through a focused hardening plan.

Explore 8 insights →
CVE analysis5 primary sources

ToolShell: recover SharePoint trust after server compromise

A defensive dossier for SharePoint ToolShell: separate patching from machine-key recovery, investigate the farm, and retire unsupported versions.

SharePoint Server
CVE-2025-53770CVE-2025-53771CVE-2025-49704CVE-2025-49706

Observed exploitation ·

Checked Nov 13, 2026Read insight
Campaign9 primary sources

Storm-2372: close the device-code trust gap

Use dated Microsoft campaign evidence to restrict device-code authentication, review device registration, and investigate token-backed access without assuming MFA alone stops the attack.

Microsoft Entra IDMicrosoft TeamsExchange Online

Observed campaign ·

Checked Nov 13, 2026Read insight
CVE analysis6 primary sources

EchoLeak: harden Copilot after CVE-2025-32711

Separate the completed Microsoft cloud fix from lasting controls for untrusted email, excessive access, Copilot grounding and investigation evidence.

Microsoft 365 Copilot
CVE-2025-32711

Vendor advisory ·

Checked Nov 13, 2026Read insight

Go deeper

The boundaries that need a closer look.

Delegated identities. Permission-aware retrieval. Agent tools. Correlation that survives scrutiny.

Explore advanced security →
  • Trace authority across tokens and services
  • Test retrieval, revocation and ingestion integrity
  • Validate tool actions and adversarial findings
  • Investigate identity changes and data access

Navigate the knowledge

Find the boundary you need to secure.

All 25 topics

Go further with specialist collections on Conditional Access, MCP, secrets, storage and detection. Explore all 25 security topics →

Read the latest

Fresh context for real security decisions.

Content updates
Agents4 min read

Govern agent connectors before they move data

Turn a Copilot Studio connector inventory into scoped data policies, verified endpoint controls, and accountable changes across environments.

Documentation checked

Primary sources, explicit review methods and dated revisions. See how the guidance is reviewed.

Inside the query library

Turn a security question
into an investigation.

KQL examples for Microsoft Sentinel, with the data prerequisites, assumptions and validation notes beside the query.

Explore 53 queries

Evidence before confidence

Know what supports the guidance.

Follow the primary sources. Check the review method. Test the control in your own environment.

Our editorial approach