DLP Configuration

What Is DLP?

Data Loss Prevention (DLP) is AgentWatch’s real-time content inspection layer. Every AI prompt and response passes through the DLP engine before reaching the LLM or the client. If sensitive data is detected, AgentWatch can block the request outright, silently redact the sensitive values, or log a warning while letting the request proceed — depending on your policy.

DLP is critical in enterprise AI deployments because employees regularly paste internal data (contracts, patient records, customer PII) into AI prompts without realizing it is being transmitted to a third-party model provider. AgentWatch catches this before the data leaves your network.

How the DLP Pipeline Works

Inbound Request (prompt text)
        │
        ▼
┌─────────────────────────────────────┐
│ 1. Built-in Regex Patterns (50+)    │
│    Fast pattern matching against    │
│    SSN, credit cards, API keys…     │
└──────────────┬──────────────────────┘
               │
               ▼
┌─────────────────────────────────────┐
│ 2. Custom Rules (if configured)     │
│    Admin-defined regex patterns     │
│    for company-specific data        │
└──────────────┬──────────────────────┘
               │
               ▼
┌─────────────────────────────────────┐
│ 3. NER Enhancement (if enabled)     │
│    Python/Presidio bridge for       │
│    contextual entity detection      │
└──────────────┬──────────────────────┘
               │
               ▼
┌─────────────────────────────────────┐
│ 4. External Webhook (if configured) │
│    POST to your own DLP service;    │
│    additional findings are merged   │
└──────────────┬──────────────────────┘
               │
               ▼
┌─────────────────────────────────────┐
│  Risk Score Calculation             │
│  (severity × category multiplier)  │
└──────────────┬──────────────────────┘
               │
        ┌──────┴──────┐
        ▼             ▼
   Score < 5      Score ≥ 5
   Allow/Warn      Block / Redact

Data Categories and What AgentWatch Detects

PII — Personally Identifiable Information

Data Type	Examples Detected
Email addresses	user@company.com
Phone numbers	US and international E.164 format
Social Security Numbers	123-45-6789
Government IDs	ITIN, EIN, passport numbers, driver’s licenses
Dates of birth	DOB: 01/15/1980
Physical addresses	Street + city + ZIP combinations
IP / MAC addresses	IPv4, IPv6, hardware MAC
Bank identifiers	Account numbers, routing numbers, IBAN, SWIFT

PCI-DSS — Payment Card Data

Data Type	Notes
Credit/debit card numbers	Visa, Mastercard, Amex, Discover — Luhn-validated
CVV / CVC codes	3–4 digit security codes
Card expiration dates	MM/YY and MM/YYYY formats
Card PINs	4–6 digit PIN patterns in context

PHI — Protected Health Information (HIPAA)

Data Type	Examples
Medical Record Numbers	MRN patterns
Provider identifiers	NPI, DEA numbers
Clinical codes	ICD-10 diagnoses, CPT procedures, NDC drug codes
Health details	Conditions, medications, lab results, vital signs
Biometric data	Fingerprint references, genetic data mentions

Credentials and Secrets

Data Type	Detection Pattern
AWS access keys	AKIA... prefix pattern
API keys	OpenAI sk-..., Anthropic sk-ant-..., Stripe sk_live_...
GitHub tokens	ghp_..., github_pat_...
JWT tokens	Three-part base64 structure
Private key headers	-----BEGIN PRIVATE KEY-----
Database URLs	postgres://, mysql:// with credentials

Risk Scoring

Each detection is scored by severity and category:

Severity	Weight
Critical	10
High	7
Medium	4
Low	1

Category multipliers: PHI × 1.5, PCI × 1.4, Credentials × 1.3, PII × 1.0

The highest-scoring detection contributes 70% of the total risk score; remaining detections contribute the balance. The result is a normalized 0–10 score.

Score Range	Action
0.0 – 2.9	Low risk — allow and log
3.0 – 4.9	Medium risk — warn
5.0 – 10.0	High / Critical risk — block or redact

Action Modes

Mode	What Happens
Block	Request is rejected with 403 Forbidden. The prompt never reaches the LLM.
Redact	Sensitive values are replaced inline with tokens like [EMAIL-REDACTED], [SSN-REDACTED], [CARD-REDACTED]. The sanitized prompt is forwarded to the LLM.
Warn	Detection is logged as a DLP event; the original prompt is forwarded unchanged.

Configuration Settings

Navigate to: Settings → General → DLP

Setting	Description	Default
DLP Enabled	Master on/off switch	Off
Scan PII	Enable PII category detection	On
Scan Financial (PCI)	Enable payment card detection	On
Scan Health (PHI)	Enable HIPAA PHI detection	On
Scan Credentials	Detect API keys, passwords, tokens	On
Action on Violation	block, redact, or warn	warn
Scan Direction	request, response, or both	both

Custom DLP Rules

Administrators can add company-specific patterns (internal project codenames, proprietary identifiers, etc.):

{
  "name": "Project Codename",
  "pattern": "\\b(PHOENIX|TITAN|NOVA)-\\d{4}\\b",
  "category": "CUSTOM",
  "severity": "high",
  "confidence": 0.9,
  "enabled": true
}

Rules are validated before saving — patterns over 500 characters are rejected to prevent ReDoS attacks.

Step-by-Step: Enabling DLP

1. Go to Settings → General → DLP.
2. Toggle DLP Enabled to On.
3. Select the categories to scan (PII, Financial, Health, Credentials).
4. Set Action on Violation:
   • Start with warn to audit detections without blocking.
   • Switch to redact once you have confirmed the false-positive rate.
   • Use block for the most sensitive categories (PHI, PCI) in regulated environments.
5. (Optional) Add Custom Rules for company-specific sensitive terms.
6. Click Save. DLP scanning begins immediately for all new requests.

Viewing DLP Events

Go to Analytics → Observability and filter by DLP Events. Each record shows:

• Timestamp and request ID
• Which categories were detected and their scores
• Whether the request was blocked or allowed
• The action taken (block / redact / warn)

DLP events are also included in the Audit Export (section 7.4).

API Reference

Method	Endpoint	Description
GET	/api/admin/company/settings	Read current DLP configuration
PUT	/api/admin/company/settings	Update DLP settings
GET	/api/admin/audit/export?type=dlp	Export DLP events