Capabilities
Key Features
Detectors that validate, not guess
Luhn-checked card numbers, mod-97 IBANs, SSA-rule social security numbers, RFC-shaped emails and E.164 phone numbers. A non-Luhn sixteen-digit string is rejected, which is why the findings survive review.
Secrets by provider signature
AWS, GitHub, Slack, Google, Stripe, OpenAI, Twilio, SendGrid, JWT and PEM signatures, with Shannon-entropy analysis as the fallback for generic tokens.
Data lineage through AI
Where sensitive data entered, which agent or assistant handled it, what it grounded, and where the output went. The path, not just the endpoints.
Classification you already own
Microsoft Purview sensitivity labels and DLP verdicts are read natively, so AI policy inherits your existing classification rather than starting a competing one.
Hidden-text and smuggling detection
Zero-width characters, Trojan-Source bidirectional overrides and tag-block smuggling — the payloads that are invisible to a human reviewer and perfectly legible to a model.
Redaction at the gateway
Where you route traffic through the AI Gateway, sensitive values are redacted inline before they reach the provider — prevention rather than a report about it.
Signal quality
A detector that cries wolf gets turned off
Most AI DLP is regex with confidence in its voice. It flags every sixteen-digit string as a card number, every high-entropy token as a secret, and within two weeks the team has muted the channel. Validation is not a nice-to-have — it is what makes the control survive contact with a real queue.
Credit cards
Luhn checksum verified. An invalid sixteen-digit sequence — an order number, a device ID — is rejected rather than reported.
Social security numbers
Validated against Social Security Administration allocation rules, so impossible area numbers are excluded.
Bank accounts
IBAN mod-97 checksum verified, which eliminates the bulk of false positives that structure-only matching produces.
API keys and secrets
Matched against known provider signatures first. Entropy analysis is the fallback for tokens with no recognisable shape, not the primary method.
Deferring to a real classifier
Where an upstream scan has already classified content with a model, the heuristic layer stands down rather than re-flagging what the classifier cleared.
Exceptions with expiry
Every mute and exception carries an owner and a hard expiry, so a suppression made during one incident does not quietly become permanent policy.
Get started with Data Security
See how Saf3AI can help secure your AI agents.