VaikoraVaikora

Vaikora › Blog › Product Updates

Classification happens at the source or it happens too late: a practical guide for security teams

Product Updates · Nov 20, 2026 · 5 min read

Three different teams in the same organization can touch the same file and assign it three different sensitivity labels. That happens because classification was added after data was created, in a downstream tool, by a process that guesses at context the original system already knew. The fix is not a better guessing algorithm. It is enforcing classification at the point of creation.

Why downstream classification fails under pressure

When a document is created in a business application, it carries context: the form it was generated from, the fields it contains, the workflow that triggered it. That context is clearest at the moment of creation and degrades as the file moves. By the time it lands in a data lake, an email attachment, or an AI agent's context window, the original structure is gone or flattened.

Classification tools that run on stored data work from text patterns, regular expressions, and statistical models. They are good at finding social security numbers and credit card patterns. They are worse at understanding that a particular spreadsheet is confidential because of what the columns represent in combination, not because any single field triggers a rule. A human who created that spreadsheet knew immediately. The downstream scanner did not.

The result: inconsistent labels, coverage gaps, and policies that cannot be enforced because there is no reliable signal to enforce against.

What "at the source" actually means operationally

Classification at the source means the system that generates or first receives the data applies a label before that data moves anywhere. Concretely:

The label becomes a property of the data, not a tag applied to a copy of it. When data moves, the label moves with it.

The enforcement dependency chain

Every downstream control depends on classification accuracy. Access control policies define what authenticated roles can read which sensitivity tiers. Data loss prevention rules fire on labeled data. Retention schedules apply by classification. Audit trails record what classification a user accessed.

If the label is wrong, all of those controls are wrong in a way that is invisible. A file marked "internal" instead of "restricted" does not trigger any alerts. No policy is violated. The access event logs cleanly. The error is silent.

This is why classification is not just a metadata problem. It is a governance foundation. Controls that look like they are working may be working correctly against the wrong labels.

Where AI agents make this more urgent

An AI agent reading enterprise data does not have judgment about what it is touching. It sees tokens. If a document contains compensation data for a named executive but carries no label, the agent has no signal to pause on. It will summarize, extract, or pass that content to a tool call based solely on whatever task it was given.

Vaikora enforces policy on AI agent actions before they execute, checking every action against deterministic rules. But Vaikora's ability to enforce a data-sensitivity policy depends on the data carrying a classification signal in the first place. When a file has no label or an inaccurate label, even inline enforcement cannot apply the right control. The agent action looks clean because the data looks clean.

Source classification closes this gap. A document classified "restricted" at creation carries that label into the agent's context window. Vaikora reads the label, matches it against policy, and can block or require approval before the agent does anything with that content.

A worked example: contract data in a retrieval workflow

A legal team uses a retrieval-augmented agent to answer questions about vendor contracts. Contracts are stored in a document management system. Some contracts contain NDA terms, some contain pricing that is confidential under third-party agreements.

If the document system classifies contracts at upload, applying a tier based on template type and detected clause categories, those labels persist into the vector index and into document chunks retrieved at query time. When the agent retrieves a chunk from a confidentiality-restricted contract, that classification signal is present in the context.

Vaikora sees the classification, compares it to the requesting user's approved data access tier, and enforces the policy before the agent responds. The enforcement decision is written to the tamper-evident audit log with the classification label, the user, the action, and the outcome.

Without source classification, the retrieved chunk has no label. The enforcement layer has nothing to check against. The control does not exist.

What to audit right now

Start with the systems that generate your most sensitive data, not the repositories that store it. For each one, check whether classification is applied at write time or at scan time. If it is scan time, map what moves between creation and scan. That interval is your uncontrolled exposure window.

Then trace one classification label forward: does it survive export to CSV, copy-paste into a new document, retrieval by an AI agent? If it does not, the enforcement built on top of that label does not apply to data that has moved.

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

More from the Vaikora blog