Security setup for AI agents

Before you enable

Two principles, both non-negotiable:

1. AI inherits security; it doesn't bypass it. If an agent can see something a user can't, your security model is broken — fix that before turning on the agent.
2. Agent actions are audit-logged as the user, not as a faceless service. If you can't tell which human approved an agent-suggested action, stop.

Step-by-step

1. Enable the integration profile option

Path: Setup and Maintenance → Manage Administrator Profile Values.

• Profile option code: ORA_ASE_SAS_INTEGRATION_ENABLED
• Set to Yes at the site level.

This unlocks the Security Console's external application integration that AI agents depend on.

2. Activate the AI Agent permission groups

Path: Tools → Security Console → Roles.

• Find the permission groups for the agents you'll roll out (each agent has its own).
• Add them to the appropriate job/abstract roles in your tenant.
• Run a role copy in a test environment first if you maintain a custom role hierarchy — never edit the Oracle-seeded role directly.

3. Scope data access

Verify each agent only sees the data its assigned users could already see:

• Same data security policies as the user role.
• Same row-level security (BU, ledger, location).
• Same record visibility rules.

If the agent appears to surface data outside the user's scope, stop and review — that's a finding, not a feature.

4. Set up audit and logging

• Confirm AI agent actions appear in the audit log with the invoking user's identity.
• Pipe the audit data into your SIEM or compliance tool the same way you do other Fusion audit feeds.
• Decide retention — most customers match their existing Fusion audit retention.

5. Pilot with a tight cohort

Don't go-live to the whole tenant. Start with 10–20 users in one location/BU, monitor for 2–4 weeks, then expand.

What to watch

• Permission creep: rolling out one agent often pulls in adjacent permissions. Diff your role definitions before and after.
• Service account drift: if you're using AI Agent Studio for custom agents, the studio may use a service principal — confirm the principal's permissions are minimal.
• Cross-tenant safety: each customer's models are partitioned in Oracle's platform. Confirm with your CSM that your tenant's AI configuration is isolated as expected.