Provided by EDB
As companies give AI agents greater autonomy—the ability to plan, make decisions, and act across systems without human approval at every step—critical questions are emerging in architecture and security reviews. When an AI agent attempts an unauthorized action, what actually stops it?
These agents run on enterprise AI models and access data across your infrastructure. Organizations are responsible for their actions, and that responsibility cannot be addressed with abstract policies that exist only on paper. AI agents do not inherently make judgment-based decisions about whether an action is appropriate. They need intelligent, context-aware rules that reflect the situation at hand.
Consider a simple rule: never open a car door. Followed literally, that rule could prevent someone from entering or leaving the vehicle. But if the car has crashed, there is a fire, or someone is injured and needs to be evacuated, the appropriate action may be the opposite. Context changes everything. If we expect AI agents to act wisely, we need intelligent, executable policies.
The instinct is to place guardrails around AI agents through instructions, policies, and oversight layered on top of the model. These mechanisms are important, but they have structural limitations. A rule such as “never open the car door” may be valid until the moment a decision must be made about whether to open it. Controls at the agent layer are reliable only when the agent’s output is predictable. Autonomy is precisely what makes that output difficult to predict.
Governance that depends on reviewing actions before they occur cannot keep pace with AI systems operating in milliseconds across multiple systems. Effective AI governance must be executable and enforced where the agent performs its work: at the operational data layer, in context, and at the moment the action takes place.
The Data Layer Is the AI Enforcement Point
AI agents create value by interacting with data. They query, retrieve, transform, and increasingly act on information stored throughout the enterprise. A policy stating that an agent must not access a specific category of data is meaningful only if the system can deny that access when the agent makes the request.
Likewise, the requirement that AI systems be auditable is practical only when organizations can reconstruct what an agent did, which data it accessed, on whose behalf it acted, and what outcomes followed. When governance controls reside at the data layer, they remain in place regardless of how an AI agent is built or how it behaves. The control becomes a property of the database rather than a promise made by the agent.
AI Agent Behavior May Be Probabilistic. Governance Cannot Be.
Organizations should not rely on an AI model to consistently follow policies. Policies must be enforced by the system itself. That is the difference between asking an actor to stay within boundaries and building boundaries that cannot be crossed.
Many of the controls required for AI agent security already exist at the data layer, including role-based and attribute-based access control, row- and column-level security, data classification and masking, policy as code, and comprehensive audit trails.
What AI agents change is not the underlying enforcement mechanism, but the identities that the system must recognize. Identity management requires treating each AI agent as a distinct principal with its own identity and declared purpose. That identity should be established at the beginning of the session and linked to the user or application the agent represents.
When purpose is tied to identity, policy engines can evaluate an agent’s purpose in the same way they evaluate roles, departments, and other attributes. Audit records can then capture not only who acted and which data they accessed, but also what the agent stated it was authorized to do.
In practice, an effective framework includes nine controls organized into three areas:
Enforce
-
Role-based and attribute-based access control applied at query time for both AI agents and users
-
Dynamic column masking enforced through the same policy path
-
AI agent identity established as a first-class principal, with its declared purpose bound to the identity and the represented user recorded at the beginning of the session
Prove Through Visibility
-
Data classification and tagging used to drive access policies
-
Session-level audit logs that record which agents acted on behalf of which users and for what declared purpose
-
End-to-end pipeline observability that traces an outcome back to the request that produced it
Unify and Strengthen
-
Centralized, portable policy management
-
Encryption for data at rest and in transit
-
Consistent policy enforcement across on-premises, cloud, sovereign, and air-gapped environments
“Declared purpose makes a difference. It becomes an attribute that the access layer already understands and evaluates through the same policy path as role-based and row-level security. The enforcement mechanism remains the same. What changes is that the agent’s purpose becomes part of what the system evaluates and what the audit record can prove later,” said Priyanka Jain, vice president of product management, data and AI governance at EDB.
No matter where an organization is in its AI adoption journey, data-layer enforcement can help accelerate deployment safely. The controls already exist in the database. The key difference is ensuring that every AI agent must pass through them.
Digital Guardrails, Not Locked Doors
The goal of AI governance is not to prevent agents from doing useful work. It is to define how far agents can go, which data they can access, what they can change, when escalation is required, and how the organization can reconstruct events if something goes wrong.
This approach makes AI agents identifiable, scoped, monitored, and auditable. Organizations can adopt agentic AI faster because security, risk, and executive teams can trust the underlying operating model.
Open, Sovereign, Sourced, and Enforced
Built on open source Postgres, this open foundation enables enterprises to maintain control over where data resides, who can access it, and which policies govern it—without giving up governance to layers they do not own or cannot inspect.
For regulated industries, the combination of data sovereignty and source-level enforcement is more than a differentiator. It is a prerequisite for deploying autonomous AI agents in production.
AI agent systems will continue to become more capable and autonomous. That is a reason to think carefully about where control resides—not a reason to slow innovation. Organizations that strengthen governance at the data layer can move forward with AI more confidently because protecting enterprise data requires more than good intentions. It requires controls that are implemented, enforced, and verifiable.
EDB Postgres AI is an open, enterprise-grade, sovereign data and AI platform that integrates transactional, analytical, and AI workloads with governance applied where the data resides. See the EDB white paper for the complete framework. Manage Agentic AI at Enterprise Speed.
Max Romanenko is EDB’s chief technology officer.
Sponsored articles are content created by companies that pay us to post or have a business relationship with VentureBeat, and are always clearly marked. For more information, please contact us [email protected].
Source: venturebeat.com