The Pulse
Oracle Puts AI Trust Controls Into Its Database Architecture
Oracle announced Deep Trust AI Architecture on October 26, a framework that puts data permissions and business rules inside Oracle AI Database. It also covers controlled data APIs, natural-language SQL and AI-generated applications built wi
AI.info Team · Written with AI · fact-checked ·
“AI models can now enable amazing productivity and innovation. However, enterprises can’t realize these benefits if they can’t trust what AI generates. AI trust is the biggest challenge we face today.”
Juan Loaiza, executive vice president, AI Database Technologies, Oracle
Oracle announced its Deep Trust AI Architecture on Monday, October 26, framing it as a way to put safeguards beneath AI applications and agents rather than relying on each one to enforce company policy correctly. The plan brings together capabilities across Oracle AI Database, Autonomous AI Database and Oracle APEX AI Application Generator. Oracle divides the architecture into four layers: software platform, data, data access and application development.
Four layers, with enforcement below the app
The announcement responds to a specific weakness in application-level controls: an AI agent that queries a database directly may bypass rules written only into the application it was meant to use. Oracle’s Deep Data Security instead lets organizations define end-user data-access policies in Oracle AI Database, which applies them regardless of which application or agent makes the request. Oracle says the database checks a user’s identity against a cryptographically signed token issued by the corporate directory.
Oracle also describes SQL Assertions as a way to enforce business rules in the database. Its example is an expense limit: if an application or agent attempts to record a charge above the employee’s authorized cap, the database rejects it. The proposal shifts enforcement to the point where records are stored, rather than asking each generated application to implement the same rule on its own.
Controlled APIs instead of unrestricted SQL
For agent access, Oracle Trusted Data APIs expose read and write operations on logical business objects, such as a complete purchase order, rather than giving an agent unrestricted SQL access to underlying tables. Oracle says the database applies update rules and integrity controls to those objects. The design aims to limit an agent’s reach to the business data and actions needed for a task.
The announcement also addresses natural-language-to-SQL systems, which can produce wrong answers when a model misunderstands a question or company terminology. Oracle Deep NL2SQL uses business topics, subtopics and database metadata to provide context for interpreting requests and selecting data. Users can flag incorrect answers; a knowledge expert, assisted by an AI agent, can update the relevant knowledge, while Knowledge Change Verification checks previously validated “gold” answers again before the change is published.
APEX puts generated apps through a runtime
For application development, Oracle’s APEX AI App Generator translates a natural-language description into APEXlang, an open, declarative application language. The APEX engine then runs that application definition, leaving security and other infrastructure functions to the engine instead of asking the AI to produce thousands of lines of custom code. Oracle says developers can use the approach through VS Code, Codex, Oracle APEX or an AI agent.
At the platform layer, Oracle points to Autonomous AI Database’s automatic patching across the database engine, operating system, Exadata storage and firmware, along with rolling updates and encryption for data at rest and in motion. The company says the service is available on leading hyperscalers and on premises. It also says security and patching tools for co-managed cloud and on-premises systems are free or highly discounted through the end of February 2027.
Oracle presents an architecture, not deployment results
Oracle calls Deep Trust a blueprint and connects existing products and capabilities under a single architecture. The announcement describes mechanisms and intended safeguards, but does not provide independent test results, customer deployment figures or comparative measurements showing how much risk the design reduces. Nor does it give separate pricing for the named capabilities.
The practical distinction is where Oracle says rules should live: permissions and business constraints in the database, agent operations through governed business objects, and generated applications inside a managed runtime. That approach makes the database the enforcement point even when an AI system supplies the request.