Oracle Hardens Database Security Against AI-Discovered Vulnerabilities and Autonomous Agents

Cutting-edge AI models are identifying software vulnerabilities at scale and exposing weaknesses in how applications enforce security rules. This shift is pushing database vendors to release larger patches more frequently, while making real application testing essential to prevent those fixes from destabilizing production environments. At the same time, AI systems are circumventing security logic that was traditionally embedded in application code.

Oracle Corp. is tackling the challenge through two complementary strategies: strengthening its database code against external threats and repositioning security controls to operate at the data layer rather than the application layer. The initiative took shape after the company obtained early access to an advanced AI model and tested it against its own codebase, according to Juan Loaiza, executive vice president of Oracle Database Technologies.

Back in March, we got early access to the Mythos model and we ran it against our code and it started finding all these security vulnerabilities. It's turning up some very serious security issues. And so we said, 'Hey, this is a code red moment. The house is on fire.'

Juan Loaiza, executive vice president of Oracle Database Technologies

Scaling patches and automating validation

Oracle's primary response to AI-discovered flaws is flowing into its quarterly security updates, which have expanded significantly in scope. A single quarterly patch released earlier this year closed more than 1,400 vulnerabilities spanning Oracle's product lineup. Customers face pressure to apply patches consistently while managing the risk that large updates could introduce regressions in mission-critical systems.

To address this tension, Oracle is leveraging Real Application Testing, a capability that captures actual production workloads and validates them against patches before deployment. According to Loaiza, this approach is necessary because enterprise applications are extraordinarily intricate, often running critical infrastructure for entire countries or industries.

We have something called Real Application Testing where we can capture real workload and run it to make sure that your workload is not affected by these patches. It's not quite as simple as just apply because the applications that people run are extraordinarily complex. They run entire countries, entire industries.

Juan Loaiza

Moving security enforcement to the data layer

Patching addresses external attackers, but internal AI deployment presents a distinct problem. Historically, applications have enforced access controls and security policies at the code level, but AI systems now generate much of that code or bypass it entirely by accessing data directly. Oracle's answer is Deep Data Security, which became available in April and works by passing an authenticated user's identity as an encrypted token down to the database layer itself.

Oracle has since expanded these authorization controls to encompass remote databases, lakehouse architectures and associated data tables. The underlying principle is that security rules must be enforced at a level where AI cannot circumvent them.

The new world as I see it is all these end-user security rules have to be enforced where AI can't bypass them, which means all the way down in the data. The AI agent has to get your identity. It's acting on behalf of you. It can't leak the data if it can't see the data.

Juan Loaiza

Protecting business logic from AI agents

Business rules follow the same vulnerability pattern. Application code has traditionally prevented invalid transactions—such as approving a credit card charge that exceeds a customer's limit—but an AI agent instructed to enforce such rules may not reliably do so. Oracle is addressing this by providing declarative mechanisms to define business rules and introducing SQL Assertions that validate whether data modifications comply with established constraints.

https://www.youtube.com/embed/05nqgG1zbfE?list=PLLTiNyEQRCms

There's credit limits, there's all sorts of things that are business rules that can't be violated. And again, if we move those all the way down into the data, then no matter how you access the data, what mechanism you use, how AI is used, it cannot violate those rules.

Juan Loaiza