Oracle event: AI agent access controls belong in the database, analysts and executives say
Oracle executives and outside analysts, in a virtual event covered by theCUBE, argued that access rules for AI agents should be enforced inside the database rather than in each application, as agent-driven attacks put pressure on enterprise resilience.
Juan Loaiza, executive vice president of Oracle Database Technologies, said prompt-level defenses are not enough because users can rephrase requests to steer an agent past them. "People are very clever about how they can phrase things in order to make AI do things that you would have basically thought you had prevented by prompting the AI," he told theCUBE. If the data layer enforces that each end user can see only their own records, he said, an agent acting on that user's behalf can reach only that user's data no matter how the request is worded.
Oracle combines enforcement at database access with centralized management through Oracle Data Safe, its cloud-based security service, and Oracle Database Security Central for self-managed deployments, according to Vipin Samar, senior vice president of database security, who appeared with David Knox, vice president of database security product management. Samar said AI moves easily and naturally from one database to another, so effective protection requires a 360-degree view across the entire fleet, across users and across data.
For partners building agent-based applications, that architecture separates the queries an agent generates from the records it is authorized to retrieve. Roger Cornejo, technology innovation principal director for generative AI and AI at Accenture Enkitec Group, said the group used Oracle Deep Data Security to enforce access rules inside the database, letting agents query a shared knowledge repository while preserving different authorization levels. "We have dozens, probably 40 or more agents that are hitting this application," Cornejo told theCUBE. He said the system has been tested to scale to 300 simultaneous queries returning most results in 15 seconds or less.
Krista Case, an analyst at theCUBE Research, said agent-fueled attacks are changing where enterprises place trust as software gains the ability to reach sensitive data and act on users' behalf. The problem runs from controlling individual requests to governing actions that span multiple systems and business processes, she said, and resilience has become part of the security conversation. "Resilience really needs to be built in up front because there is a recognition that we cannot just build the fences high enough," she said.
Holger Mueller, vice president and principal analyst at Constellation Research, said the industry spent this year arguing over which agentic framework is better. "Next year is going to be the big renaissance of scaling the backend systems," he said. According to theCUBE's coverage, delaying adoption leaves organizations exposed to AI-powered threats while technical debt compounds.
Johnnie Konstantas, group vice president of Oracle Cloud Engineering, said readiness also depends on knowing where agents are being built and which sensitive assets they can reach. That inventory, she said, helps organizations establish controls around delegated work.