Note
Access to this page requires authorization. You can try signing in or changing directories.
Access to this page requires authorization. You can try changing directories.
AI systems are now part of the core enterprise technology estate, including AI applications, agents, models, and the data and services they use. As organizations adopt AI, these systems introduce new access paths, dependencies, and risk conditions that must be governed as part of end-to-end security architecture.
In the Microsoft Security Adoption Framework (SAF), the AI pillar applies Zero Trust principles to AI workloads and operations: verify explicitly, use least privilege, and assume compromise. This guidance helps you protect AI capabilities while supporting business outcomes, operational reliability, and continuous improvement.
What this technology pillar covers
This pillar focuses on securing AI systems as workloads and business-critical resources, including:
- AI applications and systems
- AI agents and agent runtimes
- AI and agent identities
- Models, prompts, and context
- Data used for grounding, training, and inference
- Tools, plugins, connectors, and external dependencies
- Agent-to-agent and agent-to-service communications
- AI infrastructure and hosting environments
Core Zero Trust control areas for AI
Use this pillar to plan and implement controls across these areas, aligned to the implementation areas in the AI security workshop:
- Map and assess AI risk
- Discover and inventory AI applications, agents, and services.
- Prioritize risks and establish ownership, governance, and acceptable-use guardrails.
- Continuously monitor risk signals and adjust priorities as usage evolves.
- Register agents and enforce lifecycle governance
- Register agents in a centralized inventory.
- Classify agents by purpose, business criticality, and risk.
- Define ownership, publishing, certification, and lifecycle controls.
- Secure AI authentication and access
- Require strong authentication and authorization for users, services, and agents.
- Apply least privilege and delegated permissions for agent actions.
- Continuously evaluate identity, device, and risk signals before and during access.
- Secure AI network and service access
- Control how AI services are reached across network paths and service boundaries.
- Apply segmentation, filtering, and inspection for AI interactions.
- Reduce exposure to prompt injection and unauthorized access paths.
- Secure AI data access and usage
- Classify and protect data used in prompts, context, grounding, and outputs.
- Apply data loss prevention and usage controls to reduce oversharing and exfiltration risk.
- Enforce governance and compliance policies for AI data handling.
- Build and deploy agents securely
- Secure AI apps, APIs, connectors, and tool-chain dependencies.
- Validate trust boundaries and runtime controls before production use.
- Apply secure development, testing, and release controls for AI-enabled systems.
- Detect and respond for AI
- Monitor AI activity, prompt flows, access patterns, and tool usage.
- Detect misuse, anomalous behavior, and policy violations.
- Integrate AI telemetry into incident detection, investigation, and response workflows.
How this pillar connects to SAF
AI security relies on other SAF technologies and security disciplines:
- Technologies: Identity, Endpoints, Applications, Data, Infrastructure, and Network.
- Security disciplines: Development for secure engineering practices and SecOps for detection and response operations.
- Business scenarios: Adopt and secure AI.