Data Loss Prevention (DLP)

Real-time protection for sensitive data in AI interactions.

Kairro’s DLP engine brings enterprise-grade data protection directly into AI workflows. Every prompt and response is scanned before it leaves the browser, preventing sensitive data exposure, enforcing regulatory requirements, and maintaining auditability—without slowing users down.

Kairro DLP dashboard

How Kairro’s DLP Engine Works

A structured, multi-stage pipeline that evaluates every prompt in real time.

1) Prompt Scanning

Regex-based detectors

Curated set of detectors:

  • CREDIT_CARD – CRITICAL
  • API_KEY_LIKE – CRITICAL
  • US_SSN – HIGH
  • EMAIL – MEDIUM
  • IP_ADDRESS – LOW

Each match: pattern name, severity, snippet offsets, truncated text, rule ID.

2) Severity Calculation

Risk signals from matches

  • maxSeverity – highest severity
  • totalMatches
  • highOrAbove – HIGH or CRITICAL
  • critical – CRITICAL only

Feeds into the policy engine.

3) Policy Evaluation

Org-specific rules

Combines DLP severity with policies to produce ALLOW / WARN / BLOCK.

  • Scopes: org, team, identity
  • Criteria: minSeverity, allow/deny tools, domain restrictions, regex, custom severity/action
4) Response Generation

Evaluate endpoint returns

  • action (allow/warn/block)
  • riskLevel
  • reasons
  • eventId
  • DLP summary flags (embedded in the event)

Extension enforces immediately.

Why DLP Matters in AI Workflows

Generative AI can unintentionally leak sensitive data. Kairro prevents exposure before data leaves the browser.

Customer PII & regulated IDs

Credentials / keys

Internal documents & source

Proprietary logic

DLP Events & Logging

Event record

Stored in Event with identity, AI tool, action, redacted prompt/response, DLP flags, tokens, metadata.

DlpMatch records

Stored in DlpMatch with patternName, severity, offsets, ruleId, detector type (REGEX), linked to the event.

Redaction & privacy controls

Prompts/responses/snippets truncated to 256 chars; offsets/metadata only; logging detail configurable per org (eventsLoggingLevels).

Default DLP Behaviors

Safe-by-default enforcement via ensureDefaultPolicies.

Block

HIGH / CRITICAL

Hard block on high-severity detections out of the box.

Warn

MEDIUM

Guided warnings to keep users productive while reducing risk.

Prebuilt controls

High-risk channels

Paste sites, unapproved AI tools, high-risk providers, optional ChatGPT deny—ready on day one.

Org-Level DLP Tuning

Customizable elements
  • Severity thresholds (e.g., “Block on MEDIUM+”)
  • Custom regex detectors
  • Target: prompt / response / both
  • Allow/deny tools by severity
  • Scope: org, team, identity
Use cases
  • Prevent full document uploads
  • Block sensitive API keys
  • Enforce PCI/PII controls
  • Warn on customer names/emails
  • Higher risk tolerance for specific teams

DLP Analytics in the Admin UI

DLP summary flags

Per-event: isDlpEvent, blockedByDlp, maxSeverity, highOrAbove, critical, totalMatches.

Detailed match view

/v1/admin/events/:id/dlp returns full match list, patterns, severities, redacted snippets.

Aggregated insights

DLP counts, high/critical last 7 days, top patterns (up to 20), block/warn ratios, trends. Capped scans (~5000 events) for consistent performance.

DLP in Real-Time Workflows

Extension enforces at the moment of use; all decisions are logged with traceable metadata.

If block threshold met

Prompt is blocked.

If warn threshold met

User receives a warning UI.

If allowed

Prompt continues.

Why Kairro’s DLP Is Different

Purpose-built for AI

Prompt-level focus, structured snippet metadata, real-time evaluation, integrated with policy/governance.

Designed for Shadow AI

DLP fires even on unapproved tools; events enter the Shadow AI pipeline.

Enterprise-safe by design

Redacted logging, scoped policies, full audit logs, and SIEM/SOAR/incident workflow integration.

The Result

Kairro’s DLP gives organizations confidence to allow AI adoption safely.

Sensitive data stays protected

Users stay productive

Security teams stay in control

Compliance stays auditable