The image generated by AIYXL using AI
An AI agent can follow an instruction correctly and still lack permission to carry it out. Oracle’s Deep Trust AI architecture, announced October 26, addresses that separation.
The blueprint spans four layers: software platform, data, data access and application development.
Oracle describes automated platform patching, database-enforced access policies and SQL Assertions for business constraints. Its spending-limit example illustrates the approach: the database rejects an excessive charge even if an application or agent attempts it.
For access, Trusted Data APIs expose business objects such as purchase orders. Deep NL2SQL adds business definitions and metadata to query generation, with corrections checked against previously validated answers. For application development, APEXlang supplies a declarative definition executed by the APEX engine.
Sources:Oracle
AIYXL’s engineering reading is that permissions must survive changes of interface. A restriction implemented only in one application may leave another access path insufficiently controlled.
The announcement presents an architectural approach, rather than evidence that every component became generally available that day. Deployment reviews still need component availability, identity configuration and failure testing. Infrastructure controls can constrain actions; answer accuracy remains a separate evaluation task.
Sources: Oracle Database Insider