Data Loss Prevention (DLP)

Real-time scanning of MCP tool-call payloads for PII, credentials, and financial data — with automatic redaction before content reaches upstream servers.

How It Works

The DLP scanner runs inline within the ShieldAgent security pipeline on every tools/call request. Every string value — including nested JSON — is scanned for sensitive data. Findings are classified by severity and GDPR category.

When a finding is detected, the proxy can redact the sensitive value in place and forward the sanitised request, block the call entirely, or log only in shadow mode. The audit trail always records what was found, even when redacted — but never the raw secret.

Encoding Bypass Detection

Attackers sometimes encode PII in base64 or URL-encoding to bypass pattern scanners. ShieldAgent DLP includes encoding-aware scanning that automatically handles obfuscated payloads before applying detection rules.

Limitation: Pattern-based scanning is defence-in-depth, not a sole control. Binary blobs and custom serialisation formats may bypass pattern rules. Pair DLP with network-level egress controls and policy-engine deny rules for layered protection.

Detection Coverage

ShieldAgent DLP provides comprehensive sensitive data detection across 20+ categories with encoding-aware scanning. Detected data is classified by severity and GDPR category.

GDPR Classifications

Every finding carries a GDPR data classification so downstream systems (audit trail, compliance reports) can make informed decisions without re-parsing the raw type.

Class	GDPR article	Meaning
personal_data	Art. 4(1)	Any information relating to an identified or identifiable natural person
special_category	Art. 9	Health, biometric, genetic, racial/ethnic origin, political opinions, religious beliefs
financial_data	Various	Payment card data, bank account numbers, routing numbers
credential	—	Authentication secrets — API keys, tokens, private keys, passwords

Redaction Strategies

When a finding is detected, the proxy applies the configured redaction strategy. Redaction replaces the matched value in-place so the sanitised request can still be forwarded to the upstream server.

Strategy	Behavior
redact	Replace matched value with a redaction token. Request is forwarded with sanitised content.
block	Reject the entire tool call and return an error to the agent. Nothing forwarded upstream.
log_only	Record the finding in the audit trail but do not modify or block. Use in shadow mode rollout.

Redacted values in audit events are stored as partial matches only — never the full sensitive value.

API Reference

DLP findings are surfaced through the audit events API and the agents risk endpoint. All endpoints require a bearer token. See Authentication for details.

GET/tenants/:tenantId/audit-events?threatType=data_loss—List DLP events (data_loss finding type)

GET/tenants/:tenantId/audit-events?threatType=dlp_redaction—List auto-redacted events

GET/tenants/:tenantId/agents/:agentId/risk—Agent risk score — includes dlp_redaction signal weight

Example — fetch DLP events for an agent

bash

curl -s 'https://api.shieldagent.io/tenants/:tenantId/audit-events?agentId=:agentId&threatType=data_loss&limit=50' \
  -H 'Authorization: Bearer <token>'

Audit event shape (DLP finding)

json

{
  "id": "ae_...",
  "agentId": "...",
  "toolName": "write_file",
  "outcome": "blocked",
  "threatType": "data_loss",
  "dlpFindings": [
    {
      "type": "email_address",
      "severity": "medium",
      "redactedMatch": "j***@example.com",
      "gdprClassification": "personal_data"
    }
  ],
  "timestamp": "2026-04-24T10:32:00.000Z"
}

Risk Score Impact

DLP findings feed directly into the agent's Risk Scoring model. Blocked calls and auto-redacted requests both raise the agent's risk score and appear as signals in the risk dashboard and anomaly detection.