Agent Inbox is designed around scoped authority, approval gates, untrusted-input handling, tenant-aware access, and inspectable action records for AI agents working through email.
Incoming messages can contain spoofing, impersonation, malicious instructions, prompt injection, or unsafe requests. Agent Inbox evaluates those risks before permitting consequential actions.
An agent’s role, objectives, prohibited actions, accessible data, tools, and escalation requirements define its operating boundary.
Teams can choose draft-only, approval-gated, or autonomous operation according to the risk and reversibility of each workflow.
Decision records preserve the context, policies, tools, and approvals associated with an action so teams can review unexpected behavior.
Cross-inbox and organization-wide insight is permission-aware. Access is configured around the customer’s users, agents, and data model.
Security issues and data-handling questions can be reported directly to Agent Inbox. The public AI and Data Handling Policy contains processing and retention details.