Secure your agentic insider with SAF

Secure your agentic insider is a current AI security transformation theme. The name can evolve, but the underlying challenge is durable: AI agents can use identities, access data, invoke tools, communicate with services and other agents, and take actions that affect business processes. Organizations must govern those capabilities as carefully as other trusted participants in the enterprise.

The insider analogy helps communicate the combination of legitimate access and potential impact. It does not mean that an agent is a person or that every agent is malicious. It means that intended, mistaken, manipulated, or compromised agent behavior can create insider-style risk when access and actions are not bounded and monitored.

This page translates the theme into durable outcomes and execution paths. It does not create a separate control framework. Use the Microsoft Security Adoption Framework (SAF) to organize the work and canonical Microsoft guidance to implement it.

What this theme means

Secure your agentic insider means knowing which agents exist, who owns them, what identities they use, which data and tools they can access, what actions they can take, and how their behavior is monitored and constrained throughout the lifecycle.

This work connects several familiar priorities: secure AI and data together, apply least privilege, protect secrets, govern the software and model supply chain, prevent unauthorized actions, and prepare to detect and respond when an agent is manipulated, drifts from intended behavior, or is compromised.

Most of this work builds on established security foundations for identity, data, applications, development, infrastructure, posture, and operations. Agentic systems also introduce AI-specific risks involving delegated autonomy, tool use, memory, agent-to-agent interaction, and manipulation. Integrate protections for these risks into the existing foundations rather than treating agent security as a disconnected control model.

Durable outcome

The durable outcome is an environment in which agent autonomy is deliberate, bounded, observable, and accountable.

Success looks like:

  • Every production agent has a registered identity, accountable business owner, technical owner, purpose, risk classification, and lifecycle state.
  • Agent permissions are limited to the data, tools, connectors, services, and actions required for the intended task.
  • High-impact actions require stronger evidence, additional controls, or human authorization.
  • Sensitive data, prompts, context, memory, model inputs and outputs, and secrets are protected according to business risk.
  • Development teams threat-model agent workflows, validate trust boundaries, test failure modes, and secure dependencies before release.
  • Security teams can observe agent decisions and actions with enough context to investigate misuse, manipulation, drift, and compromise.
  • Response procedures can suspend identities, revoke tokens and secrets, isolate services, preserve evidence, and recover safely.
  • Governance continuously reassesses agents as models, data, tools, permissions, and business use change.

SAF work

Business scenarios

Use these SAF business scenarios to define outcomes and priorities:

Security disciplines

Coordinate the work across these SAF disciplines:

  • Strategy, integration, and governance defines acceptable use, autonomy boundaries, risk acceptance, policy, and accountability.
  • Security architecture defines agent trust boundaries, identity patterns, data flows, control points, and failure containment.
  • Access and identities governs agent, workload, service, and human identities and applies least privilege.
  • Data protects prompts, context, memory, grounding data, outputs, and downstream information use.
  • Development integrates threat modeling, secure design, dependency controls, testing, deployment, and lifecycle management.
  • Security posture inventories agents and exposure, validates policy coverage, and prioritizes gaps.
  • Infrastructure secures hosting environments, runtimes, models, tools, connectors, and supporting services.
  • SecOps monitors behavior and coordinates investigation, containment, recovery, and improvement.

Roles and collaboration

Use Security roles and responsibilities in SAF to assign accountability and decision rights.

The control model depends on collaboration among:

  • Organizational leadership and governance to define risk appetite and accountability for agent use.
  • Business and AI solution owners to define purpose, acceptable behavior, criticality, and business impact.
  • Security strategy, architecture, identity, data, privacy, and compliance teams to establish policy and reusable controls.
  • Application and product development teams to design, test, release, and maintain agent behavior safely.
  • Technical engineering and operations teams to run identity, data, platform, infrastructure, network, endpoint, and observability controls.
  • SecOps teams to detect anomalous behavior, investigate evidence, contain harmful actions, and coordinate recovery.
  • Legal, human resources, procurement, and third-party risk teams where agent actions, vendors, or data use cross organizational boundaries.

Technologies

Start with AI technology guidance, and apply connected controls across:

  • Identity for agent identities, authentication, authorization, and lifecycle.
  • Applications for APIs, plugins, connectors, tools, and runtime behavior.
  • Data for prompts, context, memory, grounding, training, inference, and outputs.
  • Infrastructure for hosting, models, compute, services, and isolation.
  • Network for service paths, segmentation, filtering, and agent-to-agent communication.
  • Endpoints where agents interact with user devices, local data, or device-bound applications.

Agent risk emerges from the interaction among these technologies. Do not rely on a single AI-specific control to manage it.

Implementation and product guidance

Use Microsoft secure Zero Trust implementation solutions to connect the business outcome and operating model to technical controls. Start with the following canonical guidance:

  1. Secure AI with Zero Trust to map agent identities, data, tools, connectors, communications, and runtime controls.
  2. AI security in the Microsoft Zero Trust Workshop to align business, security, development, and operations stakeholders.
  3. Discover AI apps and data to identify usage, applications, agents, owners, and data paths.
  4. Protect AI apps and data to apply protection to AI applications, access, and information.
  5. Govern AI for compliance to establish policy, oversight, and evidence.
  6. Use Zero Trust security to prepare for AI companions, including Microsoft Copilots for Microsoft Copilot scenarios.

Use focused implementation guidance across the agent lifecycle:

Keep detailed product configuration in canonical Microsoft documentation. Use this page to define the durable outcome, assign work through SAF, and route each team to the implementation guidance for the controls it owns.