Vaikora › Blog › Product Updates
Every policy you write is only as good as the labels your data carries when an agent reads it
Most enterprise data security failures trace back to the same gap: data leaves a system, crosses a policy boundary, or gets consumed by an automated process, and nothing attached to it says what it is or what should happen to it. Classification applied at the source closes that gap before it becomes a control problem.
Why source classification matters more than perimeter controls
Perimeter controls work by watching where data goes. Source classification works by encoding what data is before it moves. The difference is significant in practice.
A database column holding Social Security Numbers looks identical to one holding internal employee IDs until something reads the schema and applies a label. A document with salary ranges has the same file extension as a public press release. Without a label attached at creation or ingestion, every downstream system, including an AI agent, sees undifferentiated content.
When an AI agent calls a tool that queries a database, it does not inherit context from the analyst who originally created that data. It sees a response. If that response carries no classification signal, the runtime enforcement layer has nothing concrete to act on.
What "at the source" means in practice
Source classification means labeling data at the point where it enters or is created in your systems, not at the point where it exits. Concretely:
- A file uploaded to a document store gets a sensitivity label written to its metadata on ingest.
- A database column containing PII gets tagged at the schema level, so any query against that column inherits the classification.
- An API response carrying financial data includes a classification header that downstream consumers can read.
The alternative is classification at the edge, applied by a DLP tool scanning outbound traffic. That approach catches some things, but it means your internal processes, including AI agents, operated on unclassified data the entire time the data was inside your environment.
Classification at the source means every consumer, every policy engine, and every audit log sees the same label from the first read.
How an AI runtime enforcement layer uses source classification
Consider a concrete scenario. An AI agent is authorized to summarize customer support tickets for a product team. The tickets are stored in a CRM. Some tickets contain fields that have been classified as PII at the schema level. Others are classified as internal only.
When the agent calls the CRM tool, the response includes classification metadata on each field. The runtime enforcement layer, sitting inline between the agent and the tool call result, reads those labels before the agent processes the content. It applies a policy: PII fields in a summary response get redacted before they reach the agent's context window. The agent never sees the raw value.
Without source classification, the enforcement layer would need to inspect the content itself to decide what to protect. That means running pattern matching, entity recognition, or model-based detection on every response. That approach is slower, less reliable, and produces inconsistent results depending on how the data is formatted.
With source classification, the decision is deterministic. The label says what the data is. The policy says what to do with that label. The enforcement layer applies the rule.
Vaikora's inline policy engine works exactly this way: when classification metadata is present on a tool response, enforcement is a lookup against a policy rule, not a content inference. That is what makes policy decisions consistent at scale across thousands of agent actions per second.
The audit trail depends on classification too
An audit log that records "agent called CRM tool, response returned" tells you very little during an investigation. An audit log that records "agent called CRM tool, response contained fields classified PII-High and Internal-Confidential, PII-High fields were redacted per policy GDPR-002, Internal-Confidential fields were passed through" tells you exactly what happened and why.
Auditability scales with classification quality. If source labels are incomplete or inconsistent, the audit record reflects that uncertainty, and investigation becomes guesswork.
What to check in your environment
Start with the systems your AI agents query most frequently: databases, document stores, ticketing systems, communication archives. For each one, answer three questions:
- Does each data element or document carry a classification label that travels with the data, not just one that exists in a separate catalog?
- Does that label survive when the data is queried through an API or returned in a tool call response?
- Does your runtime enforcement layer read that label before passing the response to an agent, or does it only inspect content after the fact?
If the answer to any of these is no, source classification is the first thing worth fixing. Policy enforcement, runtime control, and auditability all depend on it.
See Vaikora enforce policy on your AI
Open-core AI runtime control. Self-host the MIT gateway free, or run the hosted Control Plane.
Get a demo Self-host the gateway
Vaikora