Sensitivity and data classes

A data class is a label Kratex attaches to a sensitive target: a file, an environment variable, or a host. Built-in filename classification labels the common cases: credential stores like .aws/credentials, .npmrc, and .kube/config carry credentials; key material and env files like *.pem, id_rsa*, and .env* carry secrets; wallet files like wallet.dat and Ethereum keystore/UTC--* carry wallet. Some sensitive directories, ~/.gnupg among them, carry no class at all; they are covered by the literal credential-path rules instead. Classes let a rule talk about a kind of data rather than a specific path.

Two conditions consume classes. targetClasses matches when the operation’s own target carries a class, so you can act on a kind of data wherever it lives:

{
  "id": "my-app.audit-credential-reads",
  "subject": { "source": "third-party" },
  "target": { "kind": "fs", "operation": "read" },
  "when": { "targetClasses": "credentials" },
  "effect": { "action": "audit", "severity": "high" }
}

exposed matches an operation by a caller that previously touched a classed target during the run. Reading a credential stamps the caller with the credentials exposure, and a later outbound call matches exposed (this is the exfiltration shape the built-in outflow rules catch):

{
  "id": "my-app.block-outflow-after-credential",
  "subject": { "source": "third-party" },
  "target": { "kind": "network" },
  "when": { "exposed": "credentials" },
  "effect": { "action": "block", "severity": "critical" }
}

Use "exposed": "*" to match any exposure, whatever the class.

Declaring your own

Add patterns to the sensitivity block to classify paths, environment variables, or hosts beyond the built-ins:

{
  "sensitivity": {
    "paths": [
      "secrets/**",
      { "pattern": "config/*.pem", "classes": ["credentials"] }
    ],
    "env": [
      { "pattern": "STRIPE_*", "classes": ["secrets"] }
    ],
    "hosts": [
      { "pattern": "*.internal", "classes": ["secrets"] }
    ]
  }
}

Each entry is either a bare pattern, which carries the generic sensitive class, or { "pattern": …, "classes": [...] } to assign classes of your own naming. Your classes then work in targetClasses and exposed exactly like the built-in ones. The full schema is in the policy reference.