Oracle unveils Deep Trust AI blueprint for database rules and generated apps

Oracle announced Deep Trust AI Architecture at Oracle AI World in Las Vegas on October 26, outlining how companies can put controls beneath the AI agents and applications handling their business data. The blueprint brings together Oracle AI Database, Autonomous AI Database and APEX AI Application Generator across four layers: the software platform, data, data access and application development.

The problem Oracle describes is becoming more visible as AI moves from answering questions to interacting with operational systems. An application might contain detailed privacy checks, yet an agent that connects directly to its database can take another route to the same information. Generated applications can also miss access rules. Oracle wants those protections enforced in shared infrastructure, reducing dependence on each application's implementation.

At the platform layer, Oracle points to Autonomous AI Database's automatic security updates for the database engine, operating system, storage and firmware. Rolling patching updates parts of the infrastructure in sequence while workloads continue. Encryption covers stored data and information in transit. These capabilities address the maintenance and exposure of the underlying systems on which AI applications depend.

For data permissions, Deep Data Security applies rules inside Oracle AI Database 26ai using the requesting user's identity and execution context. Policies can distinguish entire records, particular columns and individual cell values. Identity information comes through signed tokens from an organization's identity system, allowing the database to evaluate the actual user behind an agent's request. That makes a manager's access to employee information a database policy rather than a filtering step repeated across different applications.

Business restrictions have a separate role. Oracle SQL Assertions express conditions the database must enforce, such as an employee spending limit, so an invalid charge can be rejected when submitted. For agent access, the announcement describes Deep Trust Data APIs that expose controlled business objects, such as purchase orders. This gives applications a defined set of things they can read or change, with the database enforcing the relevant rules.

Oracle Deep NL2SQL addresses another source of trouble: translating a business question into the SQL instructions used to query a database. It combines company terminology and business metrics with information about how data is organized. Continuous Feedback Learning lets users flag incorrect answers for investigation and correction by a knowledge expert assisted by AI. Knowledge Change Verification rechecks previously validated answers before revised knowledge is published, aiming to catch a correction that would break other answers. These mechanisms seek better accuracy without establishing that mistakes disappear.

In application development, AI coding agents produce APEXlang, a structured description of what an application should do, and the established APEX engine runs that definition. The language is human readable, so teams can review, validate and keep definitions in version control. Oracle's approach concentrates execution and operational controls in the shared engine while giving developers an AI-generated specification to maintain. APEXlang requires APEX 26.1 or later, according to Oracle's product guidance.

October 26 marks the architecture announcement, with several underlying capabilities already available. Oracle had announced Deep Data Security's availability in AI Database 26ai in a blog dated May 1. APEX 26.2 reached general availability for on-premises deployment on October 6, adding Deep Data Security integration and more granular APEXlang imports. That release announcement described Autonomous AI Database, OCI and multicloud availability as forthcoming; it does not establish a completed rollout across those environments. The architecture announcement does not supply one general-availability date covering every component.