Skip to main content

Layered Security Architecture

Secure AI deployments use multiple independent controls. A content filter cannot replace authorization, and an approval dialog cannot correct an agent that can retrieve the wrong data.

Control Layers

LayerControlsMain question
IdentitySSO, roles, teams, lifecycle, conditional accessWho is the user or service?
Sharing and authorizationPrivate/shared entities, team access, use/edit permissionsWhat can the identity see or manage?
AgentPrompt, model, memory, skills, tools, interfaceWhat is the agent designed to do?
DataCollections, source permissions, classification, retentionWhich knowledge can influence the answer?
ToolConnections, provider scopes, function policy, validationWhich systems can the agent affect?
ApprovalRisk-based confirmation and approver contextWhich actions need a person?
ModelApproved providers, deployments, regions, evaluationsWhere and how is inference performed?
SafetyPrompt Shield, content safety, quality testsWhich harmful or unreliable behavior is reduced?
OperationsLogs, monitoring, limits, response, recoveryCan the organization detect and contain failure?

Design Principle

Enforce the important decision at more than one layer. For example, a production email agent can use a team-scoped service account, expose only read and draft functions by default, require confirmation for send, and log the final provider result. If one control is misconfigured, another still limits impact.

Review Order

When troubleshooting unexpected access or action, review the layers in this order:

  1. identity and team membership,
  2. entity sharing and permissions,
  3. agent data and tool assignments,
  4. connection and function policy,
  5. approval and execution history,
  6. external provider permissions and logs.

For infrastructure layers, continue with Reference Architecture and Network Security.