Skip to main content

DLP Scanner

The DLP (Data Loss Prevention) scanner inspects the content of every tool call before it reaches the target MCP server. If a pattern matches, the call is blocked and an alert is emitted.

Pattern library

Behavry ships with 26 built-in patterns:

CategoryExamples
PIISSN, passport numbers, driver's license
FinancialCredit card numbers (Luhn-validated), bank account numbers
CredentialsAPI keys, private keys, JWT tokens, bearer tokens
HealthcareMRN, NPI, DEA numbers, ICD codes
NetworkInternal IP ranges, MAC addresses
CloudAWS access keys, GCP service account JSON, Azure connection strings
Secrets.env file content, Docker secrets, Vault tokens

Hot-reload

Patterns are stored in the database and reloaded without restarting the proxy. Add or remove patterns from the DLP Rules dashboard tab or via the API.

Severity levels

SeverityAction
CriticalBlock immediately before OPA evaluation
HighBlock and emit DLP_VIOLATION alert
MediumAllow but log with elevated visibility

Cross-session correlation

Behavry tracks partial data across sessions — if an agent tries to exfiltrate a secret over multiple short calls (fragment reassembly attack), the correlator catches it.

Inbound scanning

The outbound DLP scanner is complemented by the Inbound Injection Scanner (AOC-1) which inspects tool results for embedded instructions before they reach agent context.

Inbound Injection (AOC-1)