DLP, Secrets Detection, and Tokenization

Use this when prompts, retrieved content, tool arguments, model outputs, or reports may contain credentials, tokens, private keys, PII, or customer-sensitive data.

Firefish DLP runs locally. It can detect sensitive patterns, return findings without raw values, redact text, and optionally tokenize values within a local session scope.

What it detects

• API-key-like values.
• Bearer-token-like values.
• Password key/value pairs.
• Private-key blocks.
• Database URLs.
• High-entropy secret-like strings.
• Emails and phone numbers.
• Luhn-valid credit-card-like numbers.
• SSN-like values.
• IP addresses.
• Custom local regex patterns supplied by settings or request metadata.

Local-first defaults

• DLP_ENABLED=true by default.
• DLP_TOKENIZATION_ENABLED=false by default.
• Raw sensitive values are not returned in findings, API metadata, or audit previews.
• Token maps are local, in-memory, and session-scoped by default.
• Detokenization fails closed unless tokenization is explicitly enabled.

Redact a synthetic secret

curl -X POST http://localhost:8000/v1/dlp/redact \
  -H "X-API-Key: change-me-local-dev-key" \
  -H "Content-Type: application/json" \
  -d '{
    "text": "Synthetic customer email analyst@example.test and fake secret FIREFISH_FAKE_SECRET_001."
  }'

Tokenize for a local session

from firefish import FirefishClient

client = FirefishClient.from_env()

tokenized = client.tokenize(
    text="Contact analyst@example.test about FIREFISH_FAKE_SECRET_001.",
    tenant_id="local-demo",
    session_id="session-001",
)

print(tokenized["tokenized_text"])

Use tokenization only when you need reversible placeholders and have an approved local token-vault operating model.

Where DLP applies

Firefish uses DLP in direct DLP APIs, scan metadata, output safety, RAG firewall analysis, chat gateway preprocessing, and audit-safe previews. The goal is to keep sensitive values out of model context and operator views unless an explicit privileged workflow requires otherwise.

Firefish does not replace a full enterprise DLP program. It is a local AI-boundary control for sensitive data that may flow through LLM applications.