Agentic Assistant security

Learn how the Agentic Assistant protects data through RBAC, approvals, regional processing, and action auditing in Cortex AgentiX,

The Agentic Assistant is built on responsible AI principles to ensure its use is safe, fair, and trustworthy. We design our AI to be transparent about its actions, accountable for its decisions, and fair in its operations, avoiding biases.

The following describes how the Agentic Assistant protects sensitive data and gives you control and understanding over its automated actions.

Access control and permissions

User roles and RBAC options

Instance and Account admins control user access to the Cortex Agentic Assistant. Cortex AgentiX uses role-based access control (RBAC) to govern chat access and permissions to view, create, edit, delete, enable, and disable agents and actions in the Agentic Assistant Hub.

Action Execution Scope

Agents can only use assigned actions. Execution is limited to the user's existing permissions in your Cortex AgentiX tenant. If a required integration is inactive, its commands and actions that wrap them do not work.

To perform actions in Slack, your Slack email must match your Cortex AgentiX user email. This ensures the system follows your assigned RBAC permissions. The system blocks actions when you lack the required permissions.

Data security and control

How sensitive data is protected

Data is hosted and encrypted by default in a dedicated Google Cloud Platform (GCP) project. Your IAM permissions isolate and protect the data. Google's multi-tenant architecture enforces strict data separation between customers.

User approval for sensitive actions

Sensitive actions require explicit user approval before execution. They never run automatically. This gives you final control over critical or data-modifying steps.

Data user policy

Prompts and outputs are processed only to generate the immediate response. They are not collected for model training or shared with third parties.

Data residency

All prompts and responses stay within the region's compute boundary.

Transparency

How the Cortex Agentic Assistant maintains transparency

You can review how the agent reaches an answer. Click the down arrow next to Plan to view the interpreted input, planned steps, and actions used. You can also view JSON artifacts created when data was retrieved or an object was created.

All agent actions are saved in an audit dataset. You can see which agent ran each action and which user invoked it.

Chat logs and Slack-initiated actions are stored in the Cortex AgentiX database. They include a Slack prefix or metadata tag.