US Database Giant Oracle Unveils 'Deep Trust' Architecture, Pushing AI Governance Down to the Database Layer
Oracle launched a new architecture at Las Vegas, advocating for built-in security and business rules at the database level to prevent AI agents from bypassing application-layer restrictions.
American database software giant Oracle announced its 'Oracle Deep Trust AI Architecture' blueprint on October 26, 2026, at the 'Oracle AI World 2026' conference in Las Vegas. The core idea of this architecture is to embed data governance, security controls, and business rules directly into the database infrastructure layer, rather than relying solely on application-level protections, aiming to fundamentally prevent AI agents from bypassing enterprise controls.
Juan Loiza, Senior Vice President of Oracle AI Database Technology, stated at the event that current AI issues are similar to those caused by "overconfident teenagers," who may misunderstand requests, leak privacy, or be exploited by malicious actors. He emphasized that simply setting guidelines for AI is insufficient because AI might ignore or break these rules. Therefore, enterprises need to build trust mechanisms deeply within each layer of the platform stack—including the software platform, data, data access, and application development—making it difficult for AI models to evade them.
The architecture includes several specific technical measures. First is 'Oracle Deep Data Security,' which allows enterprises to define end-user access rules within the database and verify identities using cryptographically signed tokens. Second is 'Oracle SQL Assertions,' a declarative mechanism that enforces business logic (such as spending limits or age restrictions) deep within the database, ensuring that any request initiated by an application or AI agent complies with the rules. Additionally, for Natural Language to SQL (NL2SQL) scenarios, Oracle introduced 'Deep NL2SQL,' which combines organization-specific business knowledge and data context to improve the accuracy of AI-generated queries, incorporating continuous feedback learning to correct errors.