Oracle wants AI agent safety rules enforced inside the database itself

The problem Oracle is pointing at

As infosecu.technews.tw reported, Oracle laid out its thinking on enterprise AI agents in a series of articles on Nov 4. Its main claim is that the goal is not endless automation. Companies need systems with firmer judgment, governance and traceability.

The specific worry is where the safety rules sit. Traditionally, security and business rules live in the application layer, the software employees use. But agents often skip the app. They can reach data through an MCP server (a standard connector that lets AI tools use outside data) or through NL2SQL, which turns plain-language requests into database queries. In those cases, checks inside the app are easy to bypass.

Rules inside the database

Oracle's answer is the Deep Trust AI Architecture. It builds security and business rules directly into the database and the wider enterprise technology stack.

Users are identified through cryptographically signed directory tokens, which are digital proofs of who someone is. With that identity attached, the system can limit what an agent does using business conditions. Oracle's example is a payment request above an employee's approval limit. The database itself would stop it.

Receipts for what agents do

Oracle extends the idea to two products. Its AI Data Platform is meant to build reusable business context from a company's own data, documents and metadata. That gives agents a consistent, controlled setting and is meant to reduce hallucinations and misjudgments.

Fusion Cloud acts as a governed place for agents to run, keeping permissions, controls and approval requirements intact. It is also meant to produce an Outcome Receipt for audits. The receipt records the authorization behind an action, the evidence used, the decision process and the actual transaction details.

What Oracle says comes next

Oracle also previewed a future AI governance framework. It would let companies monitor agent performance, output quality, usage, operating cost and drift, meaning gradual change in model behavior. It would also support testing, certification and compliance work.

The source describes a proposal and direction, not independent test results.