Redaction & DLP
Sensitive data, stopped before it leaves your boundary
The detector is the product. Wardary scans every prompt and file, redacts or blocks what's sensitive, and does it reversibly — so your team barely notices and your data never escapes.
Detection
What Wardary looks for
Node-native pattern detectors run on every prompt and uploaded file, behind a common interface so new detectors slot in cleanly.
- Social Security numbers (with Luhn-style validation)
- Credit & debit card numbers
- Email addresses
- Phone numbers
- API keys & secrets
- Custom regex patterns you define
Name and address recognition via a machine-learning detector is on the roadmap behind the same interface; low-confidence, high-risk entities fail safe today.
Please review the contract for [NAME_1] — their account [CARD_1] and SSN [SSN_1] are referenced in section 4. Send a copy to [EMAIL_1].
4 spans tokenized · provider receives placeholders only · restored in your view on reply
Per-rule actions
Redact, block, or flag — your call, rule by rule
Each policy rule carries its own action and its own raw-retention setting, so different kinds of sensitive data are handled appropriately.
Replace the matched span with a placeholder token. The value never leaves; it's restored only in your view.
Stop the prompt from being sent at all. The user gets an inline explanation and can edit and resend.
Allow the prompt but record the match for the audit trail and later policy review.
Reversible by design, leak-safe by construction
Matched spans become stable placeholder tokens before the outbound call. When the model replies, tokens are restored to real values in your view — so the conversation reads naturally while the provider only ever saw tokens.
- Placeholders are high-entropy per-request nonces
- The token↔value map is in-memory and request-scoped
- Under a no-retain rule, the real value is never persisted anywhere
- Restore refuses unknown tokens — a hallucinated placeholder can never inject a real value
Most-restrictive-wins
When spans of different retention rules overlap, the most restrictive one wins — a pure, exhaustively-tested merge decides exactly which bytes are ever stored.
Fail closed
A high-risk span below the confidence threshold is auto-redacted, never sent. We accept occasional over-redaction over any chance of a leak.
Files too
Uploads run through the same pipeline
- Parse
Read the file
PDF, DOCX, TXT, and CSV uploads are parsed to text the detectors can read.
- Scan
Detect & redact
The extracted contents flow through the exact same detection and redaction step as a typed prompt.
- Send
Egress safely
Only the redacted contents are sent; raw files stay encrypted and tenant-scoped in object storage.
Held to a measurable bar
A detector is only as good as its recall. Wardary is built against a golden evaluation corpus with explicit precision and recall targets (we're aiming for ≥99% recall on a fixed corpus) — because for a regulated buyer, “probably caught it” isn't an acceptance criterion. Perfect recall is impossible, so we fail safe and tell you so.